Collaborative short-term load forecasting with large and lightweight models under abrupt meteorological change scenarios

1 Introduction

Short-term power load forecasting is a fundamental task for power-system dispatching, unit commitment, demand response, electricity trading, and secure operation control. Its accuracy directly affects generation scheduling, reserve allocation, peak-load management, and system security. With the development of new power systems, renewable-generation variability, distributed resources, flexible loads, and electric vehicles have increased uncertainty on both the supply and demand sides. Therefore, load forecasting is no longer a routine curve-fitting problem based only on historical load; it has become a key technical link connecting renewable-energy accommodation, demand-side management, electricity-market operation, and real-time dispatching decisions.

In renewable-rich power systems, the forecasting problem is further complicated by the reshaping of net-load curves. The duck-curve problem in renewable power systems has been investigated using a multi-task iTransformer model for net-load forecasting, showing that high renewable penetration can reshape regional load profiles and that net load should be modeled through the joint relationship among photovoltaic generation, wind power, active load, and aggregated net load (Pei et al., 2025). This finding suggests that short-term forecasting models should not only learn regular demand patterns but also capture exogenous disturbances and source-load interactions under new power-system conditions.

Short-term load is influenced by historical demand, meteorological conditions, calendar attributes, economic activity, and user behavior. Among these factors, weather is one of the most significant exogenous drivers. The sensitivity of electricity and natural-gas consumption to climate variables has been quantified, indicating that temperature changes strongly affect energy demand (Sailor and Munoz, 1997). The joint use of temperature, humidity, wind speed, and other meteorological variables has also been shown to improve short-term electricity-load forecasting (Friedrich and Afshari, 2015). In summer and winter, cooling and heating loads increase the weather sensitivity of total demand. When weather conditions are stable, the load curve usually follows regular daily and weekly patterns. In contrast, abrupt warming, abrupt cooling, heavy rainfall, or hot-humid conditions can rapidly change electricity consumption behavior, leading to peak lifting, valley shifting, increased ramp rates, and intensified local fluctuations. Such weather-mutation scenarios often become high-risk periods in which forecasting errors concentrate.

Weather-sensitive load becomes especially important in summer and winter because cooling and heating equipment account for a larger proportion of total electricity consumption. When weather is mild and stable, user behavior is mainly controlled by regular work-rest schedules and calendar patterns, so the load curve is relatively smooth and periodic. When abrupt warming, abrupt cooling, strong rainfall, or hot-humid weather occurs, however, users may change air-conditioning, heating, ventilation, lighting, travel, and commercial activity within a short period. As a result, the load curve may show peak lifting, valley displacement, increased ramping rate, and stronger local volatility. These are precisely the intervals in which forecasting errors tend to concentrate.

Existing short-term load-forecasting methods include statistical models, regression models, machine-learning models, and deep-learning models. Statistical and semiparametric models are interpretable but limited in representing complex nonlinear changes and abrupt weather-response processes. In machine-learning research, random forests improve prediction robustness through ensemble learning (Breiman, 2001), gradient boosting machines enhance nonlinear fitting by iteratively modeling residuals (Friedman, 2001), and XGBoost provides scalable and regularized tree boosting (Chen and Guestrin, 2016). LightGBM further improves gradient-boosting efficiency under large-scale feature settings, while support-vector networks provide a kernel-based framework for nonlinear regression (Ke et al., 2017; Cortes and Vapnik, 1995). These methods can handle multidimensional exogenous inputs, but they still rely heavily on manually designed features and often have limited ability to model long temporal dependencies and multi-scale dynamics.

Deep-learning methods, including recurrent neural networks, long short-term memory networks, gated recurrent units, convolutional neural networks, and deep feed-forward networks, have become important tools for load forecasting. The LSTM architecture provides a basis for modeling long-term dependencies in load sequences (Hochreiter and Schmidhuber, 1997), and gated recurrent sequence modeling supports compact temporal representation for sequence prediction (Cho et al., 2014). For electricity-demand forecasting, LSTM recurrent networks have been applied to short-term residential load forecasting (Kong et al., 2019), deep neural networks have been used for building energy-load forecasting (Marino et al., 2016), and broader deep-learning frameworks have been developed for electricity-demand prediction (Bedi and Toshniwal, 2019). However, deep models require sufficient high-quality samples and are sensitive to distribution shifts. When extreme-weather or mutation samples are sparse, they may underlearn low-frequency but high-impact scenarios, producing local forecasting bias.

Recently, Transformer-based and large time-series models have provided new approaches for long-sequence forecasting. The self-attention mechanism enables long-range dependency modeling (Vaswani et al., 2017), while Temporal Fusion Transformer and Informer further improve variable selection, gating, multi-horizon prediction, and long-sequence forecasting efficiency (Lim et al., 2021; Zhou H. et al., 2021). Autoformer uses decomposition and autocorrelation to model long-term periodic patterns (Wu et al., 2021), and subsequent Transformer variants enhance frequency-domain decomposition and patch-wise time-series modeling (Zhou T. et al., 2022; Nie et al., 2023). Combining KAN-Autoformer architectures with adaptive feature extraction and residual correction has also been shown to improve multi-energy load and carbon-emission prediction under coupled energy-system conditions (Wei et al., 2026). For power-load forecasting, large time-series models can learn seasonal patterns, long-term trends, and nonlinear weather-load couplings. Nevertheless, they usually have large parameter scales, high online-update costs, and insufficient flexibility for short-lived local disturbances. In abrupt temperature-change scenarios, a large time-series model may output smooth predictions that fit the global historical distribution but underestimate local peaks or respond late to load mutations.

Lightweight models are complementary to large time-series models because they are simple, efficient, easy to deploy, and inexpensive to update online. Gradient-boosting decision trees are suitable for structured variables such as temperature-change rates, humidity-change rates, time labels, holiday indicators, and historical residuals. If a lightweight model is trained to learn residuals of a large time-series model under meteorological mutation scenarios, it can rapidly correct predictions without frequently updating the large time-series model. Attention mechanisms can further help the residual branch assign different importance to features under different weather patterns, while SHAP provides an interpretable way to analyze feature contributions (Lundberg and Lee, 2017). Therefore, large and lightweight models should be considered complementary: the large time-series model learns global load evolution, whereas the lightweight model responds to local deviations induced by weather mutations.

The key idea of collaboration is therefore not to replace the large time-series model with a small model, nor to replace the small model with a large time-series model. Instead, the forecasting task is divided according to model strengths. The large time-series model provides stable global representation and baseline forecasting, while the lightweight model focuses on residuals associated with local meteorological disturbances. Classical load-forecasting reviews and comparative studies emphasize that model evaluation should consider accuracy, robustness, and data characteristics rather than a single error indicator (Hippert et al., 2001; Taylor and McSharry, 2007). Semiparametric and multiple-linear-regression benchmark methods further show the value of interpretable exogenous-variable modeling in short-term load forecasting (Fan and Hyndman, 2012; Hong et al., 2011). Competition and refined benchmark studies also indicate that weather, calendar, and scenario information are essential for reliable forecasting under heterogeneous operating conditions (Hong et al., 2014; Charlton and Singleton, 2014). This division allows the system to avoid frequent online retraining of large-scale parameters while still maintaining a rapid response when real-time weather and recent error feedback indicate abnormal load behavior.

Based on the above analysis, this study proposes a collaborative short-term load forecasting method for meteorological mutation scenarios. The method is designed to address the degradation of forecasting accuracy and robustness when abrupt weather changes cause local load deviations that are difficult for a single model to capture. The main contributions are threefold. First, a meteorological-mutation-aware feature representation is constructed by integrating multidimensional meteorological variables, short-term change rates, comprehensive mutation intensity, scenario labels, calendar information, and historical residual information. This representation enables the forecasting model to distinguish stable weather, gradual changes, abrupt warming, abrupt cooling, and compound mutation conditions. Second, a collaborative forecasting framework is proposed by coupling a large time-series model with a lightweight residual correction model. The large time-series model learns global load evolution and nonlinear weather-load coupling, whereas the lightweight residual model rapidly compensates for local forecasting bias under meteorological mutation scenarios. Third, an attention-gating and dynamic-fusion strategy is designed to adaptively adjust the contribution of residual correction according to mutation intensity, scenario labels, and recent forecasting errors, thereby improving both forecasting accuracy and engineering adaptability.

The remainder of this paper is organized as follows. Section 2 analyzes the load-response mechanism under meteorological mutation scenarios and presents the collaborative forecasting method. Section 3 further describes the overall collaborative forecasting framework and prediction workflow from a system perspective. Section 4 reports the case study, including dataset construction, comparison models, overall accuracy, scenario-stratified errors, error distribution, ablation experiments, and online robustness analysis. Section 5 concludes the paper and discusses future research directions.

2 Materials and methods

2.1 Load-response mechanism under meteorological mutation scenarios

Short-term load is jointly affected by user production and living patterns as well as meteorological conditions. Under normal weather, load curves usually exhibit relatively stable daily, weekly, and seasonal periodicity. Under meteorological mutation scenarios, rapid changes in temperature, humidity, wind speed, precipitation, and other meteorological factors can alter user behavior and equipment operating states within a short time, causing local load mutations, peak-valley shifts, and intensified fluctuations. Therefore, analyzing the load-response mechanism under meteorological mutation scenarios is an important theoretical basis for constructing accurate short-term load forecasting models.

2.1.1 Coupling between meteorological factors and power load

Power load can be regarded as the superposition of a base-load component and a meteorology-sensitive load component. The base load is mainly determined by production schedules, residential routines, commercial activities, and calendar attributes, and therefore has relatively stable periodicity. The meteorology-sensitive load is mainly determined by the operating states of air-conditioning, heating, ventilation, dehumidification, and lighting equipment, and is strongly affected by weather conditions. 'Accordingly, the load can be expressed as shown in Equation 1:where L t is the total load at time t, L b a s e , t is the base-load component, L m e t , t is the meteorology-sensitive load component, and ϵ t is the random disturbance term. This decomposition indicates that short-term load is jointly determined by stable demand patterns and weather-sensitive electricity consumption.

L t = L b a s e , t + L m e t , t + ϵ t (1)

In different seasons, the proportion of meteorology-sensitive load varies significantly. During hot summer periods, cooling load becomes an important factor affecting load variation; during cold winter periods, electric heating load becomes more sensitive to temperature drops; and in spring and autumn, when the temperature is relatively comfortable, the load response to meteorological factors is weaker. This indicates that the influence of weather on load has clear seasonal and scenario-dependent characteristics.

Temperature is the most important meteorological variable. The relationship between load and temperature is usually nonlinear rather than a simple linear function. When temperature lies within the human comfort range, air-conditioning and heating equipment are used less frequently and the load changes smoothly. When temperature exceeds the cooling threshold or falls below the heating threshold, thermostatic equipment is switched on intensively, and load increases rapidly with temperature variation. This relationship can be expressed as shown in Equation 2:where T t is the temperature at time t, T c is the cooling threshold, T h is the heating threshold, and η c and η h are the temperature-sensitivity coefficients of cooling load and heating load, respectively.

L met , T = α c T t − T c , T t > T c 0 , T h ≤ T t ≤ T c α h T h − T t , T t < T h (2)

In addition to temperature, humidity, wind speed, precipitation, and solar radiation also influence load. Humidity changes perceived temperature; under hot and humid weather, cooling and dehumidification demand increases. Wind speed affects building heat exchange; strong wind in winter can enhance heat dissipation and increase heating load. Precipitation affects travel behavior, commercial activity, and lighting demand. Solar radiation influences building heat gain and lighting load. Therefore, meteorological load response is the joint result of multiple meteorological factors and has clear coupling and nonlinear characteristics.

Meteorological variables do not affect electricity demand as isolated factors. In actual operation, temperature changes alter the demand for cooling or heating, humidity modifies the perceived thermal load, wind speed changes building heat exchange, precipitation affects travel and commercial activities, and solar radiation changes both indoor heat gain and lighting demand. These effects may reinforce one another during abrupt weather events. For example, rapid warming accompanied by high humidity can simultaneously increase cooling and dehumidification demand, while abrupt cooling accompanied by wind and rainfall can strengthen heating and lighting load. Therefore, the load response under meteorological mutation should be described as a coupled, nonlinear, and time-lagged process rather than as a single-variable temperature response.

2.1.2 Definition and quantitative classification of meteorological mutations

A meteorological mutation is defined as a weather-change process in which meteorological variables change significantly over a short time scale and produce a notable effect on electricity-consumption behavior and load-curve morphology. Compared with ordinary meteorological fluctuations, meteorological mutations are characterized by large variation amplitude, short duration, and strong load-response intensity. In short-term load forecasting, if such scenarios cannot be effectively identified, the model may continue historical similar-day patterns and therefore underestimate or overestimate local load changes. The meteorological feature set considered in this study includes temperature, relative humidity, wind speed, precipitation, and solar radiation, The short-term change rate of any meteorological variable is defined in Equation 3:where m t denotes a meteorological variable at time t, and k denotes the lag interval. For hourly data, k can be set to 1, 3, 6, or 24 h.

r t m = m t − m t − k / k (3)

The comprehensive mutation-intensity index is constructed as shown in Equation 4:where z T, z H, z W, z P, and z R denote the standardized changes in temperature, relative humidity, wind speed, precipitation, and solar radiation, respectively. The weights ω T, ω H, ω W, ω P, and ω R represent the relative importance of different meteorological variables.

S t = ω T z T + ω H z H + ω W z W + ω P z P + ω R z R (4)

The standard deviations in the mutation-intensity index remove dimensional differences across meteorological variables, and the weight coefficients indicate their relative importance. In practical applications, the weights can be normalized and calibrated according to regional climate characteristics, load composition, and the sensitivity of local electricity consumption to different weather factors.

Based on temperature-change direction, magnitude, and comprehensive mutation intensity, six weather-change modes are identified.

It should be noted that the thresholds in Table 1 are not fixed constants. For humid and hot regions in southern China, the thresholds related to humidity and apparent temperature can be moderately reduced; for northern winter-heating regions, the weights of low-temperature and cooling indicators can be increased; and for areas with high renewable-energy penetration, the indirect influence of solar radiation and wind-speed variations on net load should also be considered. Therefore, in practical modeling, the thresholds should be calibrated according to regional climatic characteristics and load composition. The experimental settings are summarized in Table 2.

TABLE 1. Quantitative classification of meteorological change modes.

TABLE 2. Experimental settings.

In application, the scenario-classification table should be used as an operational rule rather than as a purely descriptive taxonomy. First, meteorological observations and forecast data are standardized according to the historical training set. Second, short-term change rates and the comprehensive mutation-intensity index are calculated at each prediction moment. Third, the sample is assigned to a stable, gradual-change, abrupt-change, or compound-mutation scenario according to the temperature-change criterion and the comprehensive-intensity criterion. Finally, the scenario label is fed into the residual-correction model and the dynamic-fusion layer. This procedure makes the later forecasting model explicitly aware of the weather state under which each prediction is produced.

To further reflect the influence of temperature on human comfort, a humidity-temperature index or apparent-temperature indicator can be introduced. A simplified humidity-temperature coupling index is defined in Equation 5:where T H I t is the humidity-temperature coupling index, H t is relative humidity, H 0 is the reference humidity, and η is the humidity-correction coefficient.

T H I t = T t + η H t − H 0 (5)

When the humidity-temperature coupling index rises rapidly, even if the actual temperature changes only slightly, perceived sultriness may increase obviously, thereby increasing cooling and dehumidification load. Therefore, in hot and humid regions, meteorological mutation identification should not rely only on temperature variation but should also consider humidity-temperature coupling.

Based on the above definitions, the meteorological mutation-identification process can be summarized as follows. First, load and meteorological data are time-aligned, outliers are processed, and continuous variables are standardized. Second, the changes in temperature, humidity, wind speed, precipitation, and solar radiation are calculated at 1 h, 3 h, and 24 h scales. Third, the comprehensive meteorological mutation intensity is calculated from standardized changes. Fourth, the scenario type is determined according to temperature-change direction, change magnitude, and comprehensive mutation intensity. Fifth, the scenario label is used as an important input feature for the forecasting model. Through this process, continuous meteorological variations can be transformed into quantifiable, identifiable, and trainable scenario labels. The scenario-identification result can be written as the mapping function shown in Equation 6:where C t denotes the meteorological scenario label at time t, G · is the scenario-mapping function, Δ T t 1 h and Δ T t 3 h are 1-h and 3-h temperature changes, S t is comprehensive meteorological mutation intensity, and Δ H t 3 h and Δ W t 3 h are 3-h humidity and wind-speed changes.

C t = G Δ T t 1 h , Δ T t 3 h , S t , Δ H t 3 h , Δ W t 3 h (6)

This label transforms continuous meteorological variation into discrete scenario information that can be directly used by the forecasting model, thereby enhancing the model perception of complex weather conditions.

2.2 Collaborative short-term load forecasting method

This section describes the proposed collaborative short-term load forecasting method from five aspects: feature construction, large time-series model forecasting, lightweight residual correction, dynamic fusion, and offline training with online adaptive updating. Unlike a single-model forecasting method, the proposed approach decomposes the forecasting task into two subtasks: global-pattern learning and local-bias correction. The large time-series model focuses on learning long-term dependence and overall nonlinear evolution of load sequences, whereas the lightweight model responds to local short-term deviations caused by meteorological mutations.

The motivation for this decomposition is that meteorological mutation scenarios contain both regular and irregular information. Historical load, daily periodicity, weekly periodicity, and seasonal trends form the regular part, which is suitable for a large time-series model with strong global representation capability. Short-term weather changes, scenario labels, and recent prediction residuals form the irregular local disturbance part, which is more suitable for lightweight residual learning and fast online adjustment.

2.2.1 Input feature construction and meteorological mutation representation

Input feature construction is the basis of collaborative forecasting. To enable the model to learn regular load periodicity while identifying local disturbances caused by meteorological mutations, historical load, raw meteorological variables, meteorological change rates, scenario labels, and calendar attributes are organized into a unified multisource feature representation, as shown in Equation 7:

The feature set is designed to avoid a common limitation of using only raw meteorological values. Raw temperature or humidity describes the current weather state, but it does not explicitly tell the model whether the weather is changing rapidly, whether the direction of change is warming or cooling, or whether several meteorological factors are changing simultaneously. Therefore, the input representation combines original meteorological variables, short-time variation features, comprehensive mutation intensity, scenario labels, calendar information, and historical residual information. This design enables the model to learn both load inertia and external disturbance.where x t denotes the model input sample at time t, L t − p : t denotes the historical load sequence, t is the historical window length, M t is the meteorological feature vector, D t is the calendar feature vector, C t is the scenario label, and E t denotes historical forecasting residuals or error statistics.

x t = L t − p : t − 1 , M t , D t , C t , E t (7)

The variables in this input representation respectively describe load inertia, external meteorological driving forces, electricity-use periodicity, mutation-scenario categories, and local model bias. The meteorological feature vector is defined in Equation 8:where T t, H t, W t, P t, and R t represent temperature, humidity, wind speed, precipitation, and solar radiation; Δ T t 1 h, Δ T t 3 h, Δ H t 3 h, and Δ W t 3 h represent short-term meteorological changes; S t is mutation intensity; and T H I t is the humidity-temperature index.

M t = T t , H t , W t , P t , R t , Δ T t 1 h , Δ T t 3 h , Δ H t 3 h , Δ W t 3 h , S t , THI t (8)

This meteorological feature vector improves the model perception of short-term disturbances and compound weather changes. Standardization is used to reduce the influence of different feature dimensions on training stability. In the training stage, continuous input features are transformed according to the mean and standard deviation of the training set. Meteorological scenario labels are jointly determined by comprehensive mutation intensity, short-term temperature change, and inter-day temperature deviation, and are used to distinguish stable weather, gradual warming, abrupt warming, gradual cooling, abrupt cooling, and compound mutation. In practical training, continuous variables describe disturbance intensity, whereas discrete scenario labels express weather-change patterns. Their joint input avoids relying only on raw meteorological values while ignoring change direction and change rate.

2.2.2 Large time-series forecasting module

The large time-series model is used to extract global temporal regularities from load sequences. Compared with traditional shallow models or a single recurrent neural network, a large time-series model generally has stronger long-sequence modeling capability, feature-fusion ability, and cross-scenario generalization. It can learn multi-scale variation patterns from historical load, multidimensional meteorological variables, and calendar features. The baseline forecast is generated using Equation 9:

In the collaborative framework, the large time-series model is deliberately positioned as a baseline predictor rather than as the only forecasting component. Its advantage lies in learning long-range dependence and multi-scale temporal structure from historical load sequences, but its prediction may become smooth when the current weather condition deviates from the dominant training distribution. This smoothness is beneficial for stable periods because it suppresses random noise, but it may be harmful under abrupt meteorological changes because local peaks, sharp ramps, or delayed decline cannot be fully represented by global patterns alone.where L ^ t : t + H − 1 B denotes the H-step baseline forecast generated by the large time-series model, H is the forecasting horizon, f B · is the large-time-series-model mapping function, E t − p : t is the embedded temporal representation from time t-p to t, and θ B is the parameter set of the large time-series model.

L ^ t : t + H − 1 B = f B E t − p : t ; Θ B (9)

As shown in Figure 1, stacked encoding blocks represent deep temporal modeling units such as self-attention and feed-forward networks, which can learn daily cycles, weekly cycles, seasonal trends, and meteorological-load coupling relationships. The large time-series model mainly captures the dominant regularity of load variation, but when meteorological mutation samples are sparse or local changes are abrupt, it may output over-smoothed predictions. Therefore, local residual correction by the lightweight model is necessary.

The role of the large time-series model is therefore to provide a globally consistent baseline forecast rather than to replace all local correction mechanisms. Its advantage lies in learning stable periodicity, seasonal regularity, lagged dependence, and the long-range influence of weather variables. Its limitation appears when the testing sample belongs to a sparse or rapidly changing meteorological condition that is weakly represented in the historical distribution. In such cases, the model may maintain trajectory smoothness but fail to reproduce the magnitude and timing of local peaks. This observation is the direct reason for introducing a lightweight residual branch after the large-time-series-model output.

2.2.3 Lightweight residual correction module

The lightweight model is used to learn the prediction residuals of the large time-series model under meteorological mutation scenarios. The large time-series model provides the baseline prediction, whereas the lightweight model does not directly forecast the complete load curve. Instead, it compensates for local deviations that are difficult for the large time-series model to capture:

Using the lightweight model to learn residuals rather than complete load values reduces task complexity. The residual sequence directly reflects where and when the large time-series model is biased. If residuals become larger after abrupt warming, abrupt cooling, or compound meteorological mutation, the lightweight model can associate these deviations with mutation-intensity features, short-term temperature-change rates, humidity changes, wind-speed changes, and recent error statistics. This makes the residual branch more targeted than a second full-load predictor. The residual correction is calculated using Equation 10:where r ^ t denotes the residual correction produced by the lightweight model, f s · denotes the lightweight residual learner, and θ s denotes the lightweight-model parameter set.

r ^ t = f s z t ; Θ s (10)

Compared with direct load prediction, this module only learns the mapping relationship between forecasting errors and local disturbance features, so it is more suitable for rapid correction of local bias under meteorological mutation scenarios.

The attention-gating mechanism is formulated in Equation 11:where g t denotes the attention-gating weight, σ · denotes the Sigmoid activation function, W g and b g are the gating weight matrix and bias term, ⊙ denotes element-wise multiplication, and r t g denotes the gated residual correction.

g t = σ W g z t + b g , r ^ t g = g t ⊙ r ^ t (11)

This attention-gating mechanism increases the correction strength of the lightweight model when meteorological mutation intensity or historical errors are large, and suppresses unnecessary residual compensation under stable weather.

As shown in Figure 2, the lightweight residual-correction module consists of local disturbance features, a GBDT-based nonlinear residual learner, attention gating, and residual output. The module highlights how meteorological mutation features are learned through nonlinear interactions and how key disturbance variables are reweighted by attention. It can rapidly compensate for local errors such as peak underestimation, insufficient ramping rate, and delayed load decline, making it suitable for lightweight online deployment.

2.2.4 Dynamic fusion strategy

The final prediction combines the large-time-series-model baseline and the gated lightweight residual correction. To avoid the previous ambiguity among the coefficient symbols, the dynamic fusion coefficient is uniformly denoted by α t The one-step forecasting expression is given in Equation 12:where L ^ t denotes the final collaborative forecast, L ^ t B denotes the large-time-series-model baseline forecast, r ^ t g denotes the gated residual correction, and α t denotes the dynamic fusion coefficient.

L ^ t = L ^ t B + α t r ^ t g (12)

For multi-step short-term load forecasting, the same fusion logic is written as shown in Equation 13:where Y ^ t + 1 : t + H is the final prediction vector over the forecasting horizon, Y ^ t + 1 : t + H l a r g e is the large-model baseline prediction vector, and R ^ t + 1 : t + H s m a l l is the lightweight residual-correction vector.

Y ^ t + 1 : t + H = Y ^ t + 1 : t + H l a r g e + α t R ^ t + 1 : t + H s m a l l (13)

The dynamic fusion coefficient is modeled using meteorological mutation intensity, recent residual feedback, and high-risk scenario labels, as shown in Equation 14:where β 0 , β 1 , β 2 , β 3 are fusion-weight parameters, S t denotes the comprehensive meteorological mutation-intensity index, e t − 1 is the previous forecasting residual, I · is the indicator function, C t is the current meteorological scenario label, and Ω r i s k denotes the high-risk meteorological scenario set. In this study, the high-risk set includes abrupt warming, abrupt cooling, and compound mutation scenarios.

α t = σ β 0 + β 1 S t + β 2 e t − 1 + β 3 I C t ∈ Ω r i s k (14)

The attention-gating coefficient of the lightweight residual branch is calculated using Equation 15:where g t is the attention-gating coefficient, W and b are the gating weight matrix and bias term, and e t denotes the current residual-feedback feature. The gated residual correction can be written as r ^ t g = g t r ^ t, where r ^ t is the residual correction produced by the GBDT residual learner.

g t = σ W S t , e t , C t + b (15)

As shown in Figure 3, the dynamic fusion strategy constructs context-aware weights from meteorological mutation intensity, historical errors, and scenario labels. Under stable weather, the gating weight is small and the final prediction mainly relies on the large time-series model. Under abrupt weather changes, the gating weight increases and the lightweight residual correction has a stronger influence on the final output. Compared with fixed weighted fusion, dynamic fusion can adaptively adjust the forecasting strategy according to scenario changes, improving robustness under complex meteorological conditions.

The dynamic fusion coefficient also has an engineering interpretation. A small coefficient means that the system trusts the large-time-series-model baseline more, which is suitable for stable weather and regular load patterns. A large coefficient means that the system allows the lightweight residual branch to have a stronger influence on the final output, which is suitable when mutation intensity is high or recent errors show that the large time-series model is deviating from the observed load. Thus, the fusion strategy acts as a scenario-aware switch between stable global forecasting and fast local correction.

2.2.5 Offline training and online adaptive updating

This study adopts an offline training-online adaptive updating strategy. In the offline stage, historical load data, meteorological data, and calendar data are used to train the large time-series model so that it can learn the global regularity of load variation. Meanwhile, historical prediction residuals of the large time-series model are used to construct training samples for the lightweight model, enabling it to learn the mapping relationship between meteorological mutation features and forecasting errors. The offline training objective combines forecasting error, smoothness, and residual-regularization constraints, as shown in Equation 16:where L is the total training loss, MAE and RMSE constrain forecasting error, L s m o o t h constrains curve smoothness, L r e s constrains excessive residual correction, and λ 1, λ 2, λ 3, and λ 4 are loss weights.

L = λ 1 M A E + λ 2 R M S E + λ 3 L s m o o t h + λ 4 L r e s (16)

As shown in Figure 4, the offline stage includes a historical sample base, large-time-series-model pretraining, residual-sample construction, and lightweight-model training. The online stage includes real-time data flow, rolling prediction, error monitoring, and lightweight-model updating. The dashed feedback path represents parameter tuning triggered by forecasting errors. Through this strategy, the large time-series model maintains a stable representation of global load regularity, whereas the lightweight model performs local correction according to real-time meteorological changes and prediction errors, thereby improving forecasting accuracy and robustness under complex weather conditions.

In the online stage, the system prioritizes updating the lightweight model and dynamic-fusion parameters. If the lightweight model uses GBDT-type tree models, online adaptation can be implemented through incremental training with a sliding sample window or periodic retraining, without frequently updating the main parameters of the large time-series model. When real-time errors continuously exceed a threshold or the meteorological scenario enters the high-risk mutation set, the system updates the lightweight model and dynamic-fusion parameters first, which maintains the stability of the large time-series model while improving response speed to short-term meteorological disturbances.

This offline-online division also matches engineering deployment constraints. Offline pretraining can use long historical periods and sufficient computing resources to obtain a stable global representation. Online adaptation, by contrast, must be completed within a short dispatching cycle and should avoid expensive full-parameter updating. Updating only the residual sample set, the lightweight correction model, and the fusion weights allows the forecasting system to absorb recent errors rapidly while keeping the main forecasting backbone stable. Thus, the framework balances accuracy improvement, computational cost, and operational reliability.

3 Collaborative forecasting framework and workflow

The overall collaborative short-term load forecasting framework is shown in Figure 5. In a rolling forecasting cycle, historical load, real-time meteorological observations, numerical weather prediction information, and calendar attributes first enter the data layer. The event-aware feature layer then calculates meteorological change rates, comprehensive mutation intensity, and scenario labels. These features determine whether the current sample belongs to a stable, gradual-change, abrupt-change, or compound-mutation condition before the forecasting modules are called.

After feature construction, the large time-series model generates the baseline load trajectory, the lightweight residual branch estimates the local correction according to mutation-aware features and recent residuals, and the dynamic-fusion layer produces the final forecast. Once the observed load is returned, the latest residual is stored in the feedback buffer and used for subsequent lightweight updating. Therefore, the framework description in this section emphasizes information flow, module-calling order, residual feedback, and online adaptation, while the detailed mathematical definitions of each module remain in Section 2.2.

The collaborative forecasting algorithm can be summarized as follows. Step 1 constructs historical load windows and future meteorological features. Step 2 calculates meteorological mutation intensity, scenario labels, and local disturbance features. Step 3 calls the large time-series model to generate the baseline forecast. Step 4 calls the lightweight model to estimate the gated residual correction. Step 5 obtains the final forecast through dynamic fusion. Step 6 calculates the latest error after the observed load is returned, updates the residual sample base, and updates the lightweight model when the triggering condition is satisfied.

In practical operation, these steps form a closed loop. The historical window and future weather features provide the basic input for prediction; the mutation-intensity calculation and scenario-label identification determine whether the current condition is high risk; the large time-series model produces the main load trajectory; the lightweight model estimates the residual correction; and the fusion layer combines the two outputs according to scenario risk and recent error. After the real load is observed, the residual is stored and becomes new information for subsequent online updating.

The method has practical engineering applicability. On the one hand, the large time-series model does not need to be retrained in every forecasting cycle, which reduces computational resource consumption. On the other hand, the lightweight model has a small parameter scale and fast training speed, making it suitable for minute-level or hourly online updating in dispatching systems. The dynamic-fusion mechanism automatically adjusts residual-correction strength according to meteorological risk, allowing the model to remain stable under normal weather and respond rapidly under abrupt weather changes.

4 Case study and results

To verify the effectiveness of the proposed collaborative short-term load forecasting method under meteorological mutation scenarios, this section conducts a case study from the perspectives of dataset construction, comparison-model configuration, overall forecasting accuracy, scenario-stratified errors, error distribution, ablation experiments, and online robustness. Unlike analyses that only report overall test-set metrics, this study focuses on forecasting bias, peak-tracking ability, and local residual-correction effects during meteorological mutation periods, thereby directly evaluating engineering applicability under complex weather conditions.

The case study is arranged to test the method at several levels. Overall error metrics evaluate average forecasting accuracy, scenario-stratified metrics examine whether the method remains effective under different meteorological-change modes, curve comparison checks peak-tracking and ramping behavior, error-distribution analysis evaluates stability, ablation experiments verify the contribution of each module, and robustness analysis examines whether the framework can still work when meteorological information is imperfect. This multi-angle evaluation is necessary because a method that performs well only on average may still be unreliable during high-risk mutation periods.

4.1 Dataset and experimental settings

The case-study dataset consists of anonymized operational grid load records, matched hourly meteorological data, and calendar attributes with a time resolution of 1 h. The historical load data are used to characterize load periodicity and inertia; the meteorological variables include temperature, relative humidity, wind speed, precipitation, and solar radiation; and the calendar variables include hour, weekday, workday, and holiday indicators. Meteorological mutation samples are obtained by calculating short-term meteorological change rates and the comprehensive mutation-intensity index from the matched meteorological sequence, and then assigning scenario labels such as stable weather, gradual warming, abrupt warming, gradual cooling, abrupt cooling, and compound mutation. No synthetic load values are used in the reported model comparison. The training, validation, and testing sets are divided chronologically to avoid future information leakage caused by random splitting.

Table 3 specifies the backbone and main training settings of the large time-series model used in the experiments. The model takes 1 week of hourly historical information as input and produces a 24 h short-term forecasting horizon, which matches the day-ahead and intra-day forecasting setting of the case study.

TABLE 3. Parameter settings of the large time-series model.

Figures 6, 7 present the temporal load-temperature profile and the sample distribution of meteorological mutation scenarios in the test set. The samples include not only stable weather periods but also abrupt warming, abrupt cooling, gradual-change, and compound-mutation cases, which provides a basis for evaluating model adaptability under complex weather conditions.

The mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are defined in Equations 17, 18, and 19, respectively:where MAE is the mean absolute error, N is the number of testing samples, y i is the observed load of sample i, and y ^ i is the predicted load.where RMSE is the root mean square error. Compared with MAE, RMSE gives a larger penalty to large deviations and is useful for evaluating peak-load errors.where MAPE is the mean absolute percentage error and measures relative forecasting error in percentage form across different load levels.

MAE = 1 N ∑ i = 1 N y ^ i − y i (17)

RMSE = 1 N ∑ i = 1 N y ^ i − y i 2 (18)

MAPE = 100 % N ∑ i = 1 N y ^ i − y i y i (19)

Figure 6 shows that the load curve and temperature profile exhibit clear co-movement during weather-sensitive periods. When temperature rises rapidly, the load level increases and local volatility becomes stronger, indicating that temperature-change information should be explicitly included in the forecasting features.

Figure 7 indicates that the test set contains stable weather samples as well as abrupt warming, abrupt cooling, gradual-change, and compound-mutation samples. This distribution provides a basis for evaluating model adaptability under both regular and high-risk meteorological scenarios.

Figure 8 further illustrates the nonlinear coupling between temperature and load. Higher temperature and stronger mutation intensity are associated with increased load levels and larger dispersion, which supports the need for scenario-aware residual correction.

Figure 9 presents the hourly and daily distribution of meteorological mutation intensity. Mutation events are not uniformly distributed over time; instead, they concentrate in specific hours and days, suggesting that both temporal context and mutation labels are necessary for short-term load forecasting.

The two figures jointly show why the experiment cannot be evaluated only with average error metrics. The scatter relationship indicates that high-load samples are more likely to appear when temperature and mutation intensity increase simultaneously, whereas the heat map shows that mutation intensity is clustered in certain hours and days rather than uniformly distributed. This means that the forecasting difficulty is concentrated in local time intervals. If these intervals are ignored, a model may appear accurate on the whole test set while still producing large errors at precisely the moments that matter most for dispatching.

4.2 Comparison models and evaluation metrics

To comprehensively evaluate the effectiveness of the proposed method, comparison models are selected from deep-learning, machine-learning, large-time-series-model, and residual-correction categories. LSTM is used to represent a typical recurrent neural-network method for sequence modeling; XGBoost is used to represent a structured-feature-driven gradient-boosting method; the standalone large time-series model is used to evaluate global temporal modeling capability; Large-GBDT is used to verify the contribution of residual learning after large-time-series-model prediction; and the proposed dynamic-gated collaborative model is used to evaluate the complete framework that combines large-time-series-model forecasting, attention-gated lightweight correction, and dynamic fusion. This comparison setting allows the experiment to distinguish the effects of global modeling, structured feature learning, residual correction, and adaptive gating.

The role of each comparison model is different. LSTM tests whether a conventional sequence model can learn temporal dependence from historical load. XGBoost tests whether structured weather and calendar features are sufficient when modeled by a strong tree-based learner. The standalone large time-series model tests whether global temporal representation alone can handle abrupt scenarios. Large-GBDT tests whether residual learning helps even without attention gating and dynamic fusion. The proposed model tests the full hypothesis that global forecasting, local residual correction, and scenario-aware fusion should be integrated rather than used independently. The comparison-model settings are summarized in Table 4.

TABLE 4. Comparison model settings.

4.3 Overall forecasting performance

The overall results show that the proposed method achieves the lowest MAE, RMSE, and MAPE among all compared models. Compared with LSTM, the proposed method reduces MAE from 39.6 MW to 19.6 MW, corresponding to an improvement of approximately 50.5%; RMSE decreases from 56.2 MW to 28.5 MW, and MAPE decreases from 4.41% to 2.18%. Compared with the standalone large time-series model, MAE decreases by approximately 38.4%, indicating that the lightweight residual-correction branch effectively compensates for local deviations that are smoothed by the large time-series model.

The large time-series model captures the global trend and periodicity of the load curve, but it tends to underestimate local peaks when meteorological variables change rapidly. XGBoost is sensitive to structured weather features but has limited long-sequence modeling capability, whereas LSTM responds slowly to abrupt warming and cooling. By combining global temporal representation, attention-gated residual learning, and dynamic fusion, the proposed method reduces peak underestimation and delayed response, thereby improving both accuracy and stability.

From the perspective of mechanism consistency, this result is also consistent with the preceding analysis of weather-sensitive load. The standalone large time-series model mainly follows historical periodicity and therefore produces a smoother trajectory. XGBoost captures part of the nonlinear relation between structured meteorological features and load but lacks long-range temporal memory. LSTM learns sequential dependence, yet its hidden-state updating may be insufficient when temperature rises or falls sharply within a few hours. The proposed method combines these complementary strengths by taking the large-time-series-model forecast as the global baseline and using the lightweight model to correct the residual produced by abrupt meteorological disturbance. The overall forecasting accuracy is compared in Table 5.

TABLE 5. Overall forecasting accuracy comparison.

4.4 Forecasting curves under an abrupt warming day

An abrupt warming day is selected to examine local tracking capability. During the period of rapid temperature increase and rising meteorological mutation intensity, conventional models and the standalone large time-series model underestimate the load peak to different degrees. The proposed method follows the observed load more closely and maintains a smaller deviation around the peak, showing that the residual branch can be activated effectively when the weather-mutation signal becomes stronger.

The zoomed region in Figure 10 further shows that the standalone large time-series model produces a smoother curve and responds later to the peak arrival time. LSTM and XGBoost capture part of the upward trend, but their ramping amplitude remains insufficient. The proposed method has the closest fit near the peak, indicating that attention-gated residual correction increases the local compensation weight under abrupt warming. Figure 11 also shows that temperature growth and mutation-intensity enhancement occur almost synchronously, which provides a direct trigger for residual correction.

Figure 10 uses an external zoom-in panel to show the peak-load region without covering the main curve. The dashed box and connecting lines correspond to the local interval in the main plot, making it easier to compare the peak arrival time and ramping amplitude of different models. The curve comparison indicates that the proposed method not only reduces pointwise error but also improves the shape consistency of the predicted load trajectory during abrupt warming. Combined with Figure 11, the synchronous increase in temperature and mutation intensity provides a direct meteorological trigger for the lightweight residual branch, explaining why the collaborative model can respond more quickly than the standalone large time-series model.

4.5 Scenario-stratified error analysis

Overall test-set metrics can hide error concentration in mutation samples. Therefore, errors are calculated under stable weather, gradual changes, abrupt warming, abrupt cooling, and compound mutation scenarios. Under stable weather, the differences among models are relatively limited because regular load periodicity dominates. In high-risk scenarios, especially abrupt warming, abrupt cooling, and compound mutation, model differences become much larger.

Compared with the standalone large time-series model, the proposed method reduces MAE from 36 MW to 22 MW under abrupt warming, corresponding to an improvement of approximately 38.9%, and from 40 MW to 25 MW under compound mutation, corresponding to an improvement of approximately 37.5%. RMSE and MAPE show similar trends, confirming that the main advantage of the proposed method comes from suppressing local errors in high-risk meteorological scenarios rather than only improving average performance under stable samples.

These scenario-stratified results are important because meteorological mutation samples usually account for a smaller proportion of the whole dataset but carry higher operational risk. Under stable weather, most models can rely on load periodicity and obtain acceptable errors, so the improvement margin is naturally limited. Under abrupt warming, abrupt cooling, and compound mutation, however, load changes become steeper and less stationary. The proposed model gives larger benefits in these high-risk scenarios, showing that its improvement is produced by targeted residual compensation rather than by uniformly shifting the prediction curve.

Figures 12–14 further show that the performance advantage is scenario dependent. In stable-weather samples, most models can follow the regular daily load pattern, so the improvement margin is relatively small. In contrast, abrupt warming, abrupt cooling, and compound mutation samples involve stronger load nonlinearity and larger local residuals, causing the error of single models to increase sharply. The proposed method reduces this error concentration by activating the residual-correction branch when the mutation label and intensity indicate high risk. This confirms that the collaborative framework is not merely improving average accuracy; its main value lies in reducing prediction risk in high-impact meteorological scenarios.

4.6 Error distribution and stability analysis

The error distribution provides additional evidence for forecasting stability. The proposed model has a narrower distribution, a median closer to zero, and shorter tails than LSTM, XGBoost, and the standalone large time-series model. This indicates that the proposed method not only reduces average error but also decreases systematic bias and the probability of large-error samples.

The empirical distribution of absolute errors shifts leftward for the proposed method. Under the same error threshold, a larger proportion of testing samples is covered by the proposed model, which means that its prediction risk is lower in practical dispatching. The improvement is particularly visible in the medium-error interval, demonstrating that dynamic fusion and residual correction can stabilize predictions across heterogeneous weather conditions.

This distributional improvement has direct engineering meaning. In day-ahead or intra-day dispatching, the tail of the error distribution is often more critical than the mean value because large underestimation may cause insufficient reserve, while large overestimation may reduce economic efficiency. The more concentrated error distribution indicates that attention-gated residual correction can suppress occasional large deviations caused by abrupt weather changes. Therefore, the proposed method improves both numerical accuracy and the reliability of the predicted load trajectory.

The violin plot in Figure 15 indicates that the proposed model produces a more concentrated error distribution, with fewer extreme deviations in the upper tail. This is important for dispatching applications because a small number of large errors can lead to insufficient reserve allocation or inaccurate peak-load management. Figure 16 gives a complementary cumulative view: for the same absolute-error threshold, the proposed method covers a larger proportion of samples. Therefore, the collaborative model improves not only deterministic accuracy but also forecasting stability and operational risk control.

4.7 Ablation experiment and feature-contribution analysis

Ablation experiments are conducted by removing meteorological features, scenario labels, attention gating, and online updating. Removing meteorological features leads to the largest MAE increase, demonstrating that multidimensional weather inputs are the basis for identifying mutation-driven load response. Removing scenario labels or attention gating weakens the ability of the model to adapt residual-correction strength to different weather patterns.

Feature-contribution analysis shows that the 3-h temperature-change rate, comprehensive mutation intensity, and temperature itself contribute the most to residual correction, and these variables account for more than half of the overall contribution. This result is consistent with the mechanism analysis in Section 2, where temperature mutation and compound meteorological disturbance were identified as dominant drivers of local load fluctuations.

The ablation and feature-contribution results also verify the necessity of the theoretical variables defined in Section 2. If meteorological features are removed, the model cannot distinguish whether a residual originates from ordinary random fluctuation or from a weather-driven load response. If scenario labels are removed, samples from gradual warming, abrupt warming, cooling, and compound mutation are mixed together, weakening the selectivity of correction. If attention gating is removed, the residual branch lacks a mechanism for dynamically increasing correction strength during high-risk conditions. These results demonstrate that the proposed indicators are not decorative variables but functional inputs for improving mutation-period prediction.

Figure 17 shows that removing meteorological features causes the largest degradation, which verifies that multidimensional weather information is the foundation of mutation-aware forecasting. Removing scenario labels weakens the ability to distinguish different weather-change modes, while removing attention gating makes residual correction less selective and may lead to insufficient correction during mutation periods or unnecessary compensation under stable conditions. Figure 18 further explains the feature mechanism: short-term temperature-change rate, comprehensive mutation intensity, and current temperature dominate residual correction, which is consistent with the theoretical analysis that temperature mutation and compound meteorological disturbance are the main drivers of local load fluctuation.

4.8 Online updating and robustness analysis

In practical dispatching environments, weather-forecast errors, missing meteorological data, and load-distribution drift may reduce prediction accuracy. During rolling prediction after a meteorological mutation, the model without online updating shows a sustained error increase, whereas the model with online updating recovers more quickly and fluctuates less. This indicates that the lightweight residual model can use recent residual samples to adjust local correction relationships in real time.

When the missing ratio of meteorological data increases from 0% to 30%, the errors of all models increase, but the proposed method has a lower error-growth slope. This suggests that dynamic fusion and residual correction provide a degree of robustness when external meteorological information is degraded. Therefore, the proposed framework is more suitable for online load-forecasting applications where data quality and weather conditions may change continuously.

Robustness to imperfect meteorological information is especially important for practical application because real-time forecasting systems may encounter delayed observations, missing sensor values, and deviations in numerical weather forecasts. The proposed framework does not eliminate the effect of data degradation, but it reduces its impact by combining historical load inertia, global temporal representation, and residual feedback. When external weather information becomes incomplete, the large time-series model still provides a stable baseline, and the lightweight model can partially recalibrate the residual relation through recent errors. This explains why the proposed method has a lower error-growth rate under increasing missing-data ratios.

Figure 19 demonstrates the engineering value of online updating. After a meteorological mutation occurs, the model without online updating has difficulty adapting to the new residual pattern, so the error remains high for a longer period. With online updating, the lightweight model absorbs recent residual feedback and adjusts the local correction relationship, resulting in faster error recovery. Figure 20 shows that when meteorological inputs become incomplete, all methods deteriorate, but the proposed model deteriorates more slowly. This indicates that dynamic fusion and residual learning provide a buffer against external data degradation, which is valuable in real dispatching environments where sensor faults, forecast deviations, and data delays are unavoidable.

5 Conclusion

This study proposed a collaborative short-term load forecasting framework for meteorological mutation scenarios. The large time-series model is used to learn global load evolution, daily and weekly periodicity, and nonlinear weather-load coupling, while the lightweight residual correction model compensates for local forecasting deviations caused by abrupt warming, abrupt cooling, and compound meteorological changes. By introducing attention gating and dynamic fusion, the proposed framework adaptively adjusts the contribution of residual correction according to mutation intensity, scenario labels, and recent forecasting errors. Experimental results show that the proposed method achieves higher forecasting accuracy than the comparison models and performs better in high-risk meteorological mutation scenarios, especially in peak tracking and local fluctuation correction. The online adaptive updating strategy further improves practical applicability by updating the lightweight correction module without frequently retraining the large time-series model. Overall, the proposed framework provides an effective and engineering-oriented solution for robust short-term load forecasting under complex weather conditions. Future work will consider more extreme weather events, higher temporal-resolution data, multi-regional collaborative forecasting, and uncertainty-aware probabilistic forecasting.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Jiaqi Shi, Shenyang Institute of Engineering, China

Reviewed by: Liya Ma, University of Macau, China

Weiye Song, North China Electric Power University, China