A 30-year synthetic climatological normal dataset for Ghana generated using the CDSim framework

1 Introduction

Long-term climate datasets are fundamental to hydrological modeling, environmental simulation, climate-impact assessment, agricultural planning, and the development of data-driven forecasting systems (Nicholson, 2013, 2018; Todzo et al., 2020). Many modern analytical approaches, including machine learning and statistical downscaling, require temporally continuous and internally consistent climate records spanning multiple decades. Despite the growing availability of climate products, access to harmonized long-term datasets remains a major challenge in many regions, particularly across parts of Africa where observational station coverage may be spatially sparse, incomplete, inconsistent, or subject to licensing restrictions (Ezekiel et al., 2025; Haghtalab et al., 2019; Salami et al., 2025).

Several large-scale climate reanalysis and gridded observational products are currently available, including ERA5-Land, CHIRPS, CRU, and MERRA-2. These products have significantly improved access to regional climate information and are widely used for climatological analysis and environmental modeling (Klutse et al., 2021, 2024; Nkrumah et al., 2023). Among these, European Center for Medium-Range Weather Forecasts ERA5-Land has become one of the most widely adopted reanalysis datasets due to its high spatial and temporal resolution and global coverage. However, reanalysis products are fundamentally designed to reconstruct historical atmospheric states rather than provide fully synthetic realizations suitable for controlled stochastic experimentation, uncertainty propagation, or simulation benchmarking.

Synthetic climate datasets serve a different scientific purpose from observational or reanalysis products. Rather than reproducing specific historical events, stochastic climate simulations aim to generate climatologically plausible realizations that preserve essential statistical and physical properties of climate systems, including seasonality, temporal persistence, variability, skewness, and inter-variable relationships. Such datasets are particularly valuable for sensitivity analysis, stress testing, benchmarking of machine learning models, Monte Carlo experimentation, and educational applications requiring unrestricted and reproducible data.

Weather generators and stochastic climate simulation frameworks have long been used to model climatic variability (Jones and Thornton, 2013; Richardson, 1981; Wilks, 1999; Wilks and Wilby, 1999). These approaches typically combine probabilistic rainfall occurrence models, stochastic rainfall intensity distributions, autoregressive temperature processes, and seasonal forcing functions to reproduce realistic climate dynamics. More recent developments have extended these frameworks to incorporate non-stationary trends, multivariate coupling, and physically informed constraints. Nevertheless, openly available synthetic climatological datasets specifically developed for Ghana remain limited.

This study presents a synthetic 30-year climatological dataset for Ghana generated using the CDSim framework (Osei et al., 2026). The dataset covers the World Meteorological Organization (WMO) climatological normal period of 1996–2025 and includes monthly precipitation, minimum temperature, and maximum temperature for 47 stations distributed across Ghana. The framework integrates stochastic weather generation, autoregressive temporal dependence, seasonal structure, rainfall occurrence modeling, and physically motivated coupling between rainfall and temperature. Unlike observational archives or reanalysis products, the dataset does not attempt to reconstruct historical weather events; instead, it provides statistically plausible synthetic realizations designed for controlled experimentation and reproducible methodological evaluation.

A key contribution of this work is the integration of multivariate monthly Z-score bias correction using ERA5-Land reanalysis as an external climatological reference. The correction procedure was applied independently to precipitation, minimum temperature, and maximum temperature to improve climatological alignment, reduce systematic bias, and restore physically consistent synchronization between rainfall and temperature dynamics. The resulting corrected dataset preserves the flexibility and reproducibility of synthetic climate simulation while exhibiting strong agreement with regional climatological behavior represented by ERA5-Land.

The dataset is intended primarily for methodological development, benchmarking, simulation-based experimentation, uncertainty analysis, and educational applications rather than operational forecasting or policy decision-making. By providing a synthetic, reproducible, and openly reusable climatological product, this work contributes an accessible resource for climate and environmental research workflows in data-constrained settings.

2 Value of the data

• The dataset provides a fully synthetic 30-year climatological normal dataset for Ghana covering the period 1996–2025, consistent with the World Meteorological Organization definition of a climatological normal.
• The dataset contains no missing values and preserves physically consistent relationships between rainfall, minimum temperature, and maximum temperature, making it suitable for simulation-based environmental modeling workflows.
• Application of multivariate monthly Z-score bias correction substantially improves agreement with ERA5-Land climatology, reducing systematic bias while preserving seasonal variability, temporal persistence, and rainfall intermittency.
• The dataset offers an unrestricted and reproducible alternative for benchmarking climate-sensitive analytical methods, including machine learning models, hydrological simulations, statistical downscaling, and uncertainty propagation experiments.
• Unlike observational datasets or reanalysis products, the synthetic structure allows controlled experimentation without reproducing confidential or station-specific historical observations.
• The dataset supports sensitivity analysis, Monte Carlo experimentation, and stress testing of climate-dependent models under statistically realistic but fully reproducible conditions.
• Both raw simulated variables and bias-corrected variables are provided, enabling comparative evaluation of stochastic simulation performance and post-processing correction methods.
• The dataset is distributed in CSV and NetCDF formats, ensuring compatibility with R, Python, MATLAB, GIS software, and climate analysis workflows.
• The accompanying CDSim simulation framework enables complete reproducibility of the dataset generation process using identical parameters and random seeds.
• The dataset provides a useful educational resource for teaching stochastic climate simulation, climatological analysis, time-series modeling, bias correction, and reproducible environmental data science workflows.

3 Data description

The dataset consists of monthly synthetic climate records generated for 47 stations distributed across Ghana for the climatological normal period 1996–2025. The data were produced using the CDSim stochastic climate simulation framework and subsequently adjusted using multivariate monthly Z-score bias correction based on ERA5-Land climatological statistics.

The final released dataset (Version 3) includes both the original simulated variables and the bias-corrected variables to support reproducibility, methodological comparison, and benchmarking applications. Each record contains the following variables as shown in Table 1.

Table 1. Variables in the dataset.

Source: Authors' own compilation.

The corrected variables were generated using monthly multivariate Z-score bias correction applied independently to precipitation, minimum temperature, and maximum temperature using ERA5-Land monthly climatological statistics as the external reference distribution. For precipitation, zero-rainfall months were preserved during correction to avoid the generation of artificial rainfall during dry periods.

The dataset is distributed in both CSV (comma-separated values) and NetCDF (network common data form) formats. The CSV files facilitate tabular statistical analysis and machine learning workflows, while the NetCDF version supports spatio-temporal analysis and interoperability with GIS and climate analysis software. Metadata and variable naming conventions follow climate and forecast (CF) standards where applicable.

In total, the dataset contains 16,920 monthly observations corresponding to 47 stations over 30 years (47 × 30 × 12). The dataset is designed for reproducible climate experimentation, methodological benchmarking, and educational use rather than operational forecasting or reconstruction of specific historical events.

4 Experimental design, materials, and methods

4.1 Synthetic climate simulation framework

The dataset was generated using the CDSim framework (Osei et al., 2026) (version 0.1.2) implemented in the R statistical computing environment (R Core Team, 2025). CDSim was developed to generate physically plausible and statistically reproducible synthetic climate time series suitable for methodological experimentation, benchmarking, and simulation-based environmental research.

The simulation framework draws conceptual foundations from classical stochastic weather generators (Jones and Thornton, 2013; Richardson, 1981; Wilks, 1999; Wilks and Wilby, 1999), while incorporating additional mechanisms to preserve inter-variable physical consistency and long-term climatological structure. Unlike deterministic climate reconstruction approaches, the objective of CDSim is not to reproduce exact historical observations, but rather to generate synthetic realizations that preserve the dominant statistical and physical characteristics of tropical climate systems.

Monthly climate sequences were generated for 47 stations across Ghana covering the climatological normal period 1996–2025. The simulation framework was parameterized to reproduce generalized regional climatological behavior while avoiding replication of proprietary observational records.

4.1.1 Temperature simulation

Temperature generation incorporated three primary components:

• seasonal cyclic variability,
• long-term trend structure,
• and stochastic temporal persistence.

Seasonality was represented using sinusoidal harmonic functions to reproduce annual warm–cool cycles. Temporal persistence was introduced using an autoregressive AR(1) process to preserve climatic memory between consecutive months.

Minimum temperature (Avg.Tn) was generated using a bounded stochastic process based on truncated normal distributions to ensure physically realistic lower and upper limits. Maximum temperature (Avg.Tx) was subsequently generated using an asymmetric perturbation mechanism designed to preserve realistic diurnal temperature structure and non-Gaussian variability commonly observed in tropical climates.

To maintain internal physical consistency, the framework enforced the constraint:

Avg.Tx > Avg.Tn

for all generated records.

4.1.2 Rainfall simulation

Rainfall generation followed a two-stage stochastic process widely used in weather generator literature.

Rainfall occurrence was simulated using a first-order Markov chain to preserve wet–dry persistence and intermittent rainfall structure. Conditional rainfall intensity during wet months was then generated using gamma-distributed stochastic variability with seasonally varying parameters.

The rainfall simulation framework preserves several key precipitation characteristics:

• non-negativity,
• positive skewness,
• seasonal wet–dry cycles,
• and intermittent occurrence behavior.

To improve physical realism, a coupling mechanism was introduced between rainfall and temperature whereby increased rainfall occurrence suppresses maximum temperature through cloud-related cooling effects. This mechanism was designed to preserve the expected negative association between precipitation and surface temperature commonly observed in tropical climates.

4.1.3 Internal physical constraints

Post-processing constraints were applied to ensure climatological plausibility and numerical consistency across all generated variables. These included:

• enforcement of non-negative rainfall values,
• preservation of realistic temperature bounds,
• maintenance of Avg.Tx > Avg.Tn relationships,
• and preservation of seasonal wet–dry and warm–cool cycles.

The resulting dataset therefore represents a synthetic but physically coherent climatological system suitable for simulation-based experimentation and benchmarking applications.

4.2 External reference data

To evaluate the regional climatological consistency of the synthetic simulations, ERA5-Land reanalysis data produced by the European center for medium-range weather forecasts were used as the external reference dataset.

ERA5-Land monthly products were accessed and processed in R using the ecmwfr package (Hufkens et al., 2019) for data retrieval and the terra package (Hijmans et al., 2026) for raster extraction and spatial processing.

The ERA5-Land reference variables included:

• monthly mean near-surface air temperature,
• and monthly total precipitation.

These variables were aggregated and spatially matched to the synthetic station locations to enable climatological comparison between the simulated and reference datasets.

4.3 Multivariate Z-score bias correction

To improve agreement between the synthetic simulations and regional climatological conditions, a monthly multivariate Z-score bias correction procedure was applied using ERA5-Land reanalysis as the reference dataset. The correction was independently applied to precipitation, maximum temperature, and minimum temperature for each calendar month across the full 1996–2025 simulation period.

Following standard climatological bias-correction approaches (Maraun et al., 2017; Wilks, 2019), the corrected variable was computed as:

Xcorr=μERA5+(Xsim-μsimσsim)σERA5

where X_*sim* represents the simulated variable, while μ and σ denote the monthly mean and standard deviation of the simulated and ERA5-Land reference series for the corresponding calendar month.

The correction was applied separately for precipitation, minimum temperature, and maximum temperature. For precipitation, an occurrence-preserving constraint was imposed such that months with zero simulated rainfall retained zero values after correction. This prevented the generation of artificial rainfall during dry periods while preserving rainfall intermittency and wet–dry occurrence structure.

The correction framework was designed to improve climatological agreement and inter-variable physical consistency between the synthetic simulations and ERA5-Land reference climatology. In particular, the corrected series better reproduced the cooling behavior associated with peak rainfall periods, thereby improving inter-variable physical consistency within the synthetic climatological system.

4.4 Technical and statistical validation

As the dataset is synthetic, validation primarily focuses on whether the simulated data reproduce expected climatological behavior and regional climatological consistency rather than reconstructing exact historical observations.

Validation was conducted using the internal diagnostic function validate_climate_internal() in CDSim, which evaluates multiple aspects of the generated data:

Physical consistency: all records satisfy key constraints, including Avg.Tx > Avg.Tn and non-negative rainfall values.

Statistical properties: summary statistics (mean, variance) and distributional characteristics are consistent with expected climatic ranges. Rainfall exhibits positive skewness, as commonly observed in precipitation data.

Temporal persistence: Lag-1 autocorrelation in temperature and rainfall confirms realistic climatic memory, aligning with stochastic weather generator behavior (Wilks, 1999).

Cross-variable relationships: a negative correlation between rainfall and maximum temperature is observed, reflecting the cooling influence of cloud cover.

Trend behavior: simulated series reflect imposed long-term trends (e.g., gradual warming or drying), consistent with the parameterization.

Seasonal structure: monthly aggregation reveals clear seasonal cycles, including identifiable wet and dry periods, broadly consistent with patterns reported for comparable climatic regions (Nicholson, 2013, 2018; Nicholson et al., 2018).

Overall, the dataset demonstrates internally consistent and climatologically plausible behavior, supporting its use for methodological testing, simulation studies, and teaching applications.

The validation metrics (Table 2) confirm that the simulated dataset satisfies physical constraints, reproduces realistic statistical properties (e.g., rainfall skewness and variability), and captures expected temporal persistence and seasonal patterns. The negative correlation between rainfall and maximum temperature further supports the physical consistency of the generated data.

Table 2. Internal physical and statistical consistency diagnostics for the raw synthetic simulations.

Tmin, minimum temperature (Avg.Tn); Tmax, maximum temperature (Avg.Tx); SD, standard deviation; Corr, Pearson correlation coefficient; ACF(1), lag-1 autocorrelation coefficient; Rain, total monthly rainfall (Sum.Rf). Source: Authors' own compilation.

Table 3 indicates that the multivariate Z-score correction substantially improved agreement between the CDSim simulations and ERA5-Land climatology. Corrected temperature exhibited very strong correspondence with ERA5-Land (r = 0.96), negligible mean bias, and low RMSE values, indicating effective removal of systematic temperature bias. Rainfall also showed substantial improvement in climatological agreement (r = 0.86), although residual bias and distributional differences remained, reflecting the inherent intermittency and nonlinear behavior of precipitation processes.

Table 3. External validation statistics comparing corrected CDSim outputs against ERA5-Land.

RMSE, root mean square error; MAE, mean absolute error; KS, Kolmogorov-Smirnov statistic. Source: Authors' own compilation.

Figure 1 depicts the methodological workflow adopted in generating the synthetic climatological dataset.

Figures 2–5 present comparisons between the raw synthetic simulations, bias-corrected synthetic simulations, and ERA5-Land reference climatology for temperature and rainfall variables. Figure 6 illustrates the monthly synchronization between rainfall, minimum temperature, and maximum temperature for the Abetifi station across the 1996–2025 climatological normal period. All figures were generated in R using the ggplot2 package (Wickham, 2016).

5 Usage notes and limitations

This dataset is not intended for operational forecasting or policy decision-making. Instead, appropriate applications include:

• Testing climate-sensitive machine learning models (e.g., LSTM, DeepAR).
• Benchmarking statistical downscaling methods.
• Educational demonstrations of climate variability.
• Prototyping environmental modeling workflows prior to data licensing.

Limitations include:

• No explicit representation of observed historical extreme events (e.g., droughts or floods).
• Stochastic variability does not correspond to specific observed years.
• The dataset represents generalized regional climate patterns and does not capture fine-scale local variability.

6 Conclusion

This paper presents a 30-year synthetic climatological dataset for Ghana generated using the CDSim stochastic climate simulation framework. The dataset was developed to provide a reproducible and openly reusable climate resource suitable for methodological experimentation, benchmarking, uncertainty analysis, and educational applications. Covering the 1996–2025 World Meteorological Organization climatological normal period, the dataset includes monthly precipitation, minimum temperature, and maximum temperature generated for stations distributed across Ghana.

The simulation framework integrates seasonal forcing, autoregressive temporal persistence, stochastic rainfall occurrence and intensity modeling, and physically motivated coupling between rainfall and temperature. Internal diagnostic evaluation demonstrated that the generated climate series preserve important climatological characteristics, including realistic seasonal variability, positive rainfall skewness, temporal persistence, and physically consistent relationships between precipitation and temperature.

To improve regional climatological consistency, a multivariate monthly Z-score bias correction procedure was applied using ERA5-Land reanalysis as the external reference dataset. External validation showed strong agreement between the corrected simulations and ERA5-Land climatology, particularly for temperature dynamics, where systematic bias was substantially reduced and distributional similarity was preserved. Rainfall simulations also showed marked improvement after correction, although residual differences remained, reflecting the intermittent and highly nonlinear nature of precipitation processes.

Unlike observational or reanalysis datasets, the synthetic structure of the dataset enables controlled experimentation without reproducing station-specific historical observations. The dataset is therefore particularly suitable for simulation-based environmental modeling, machine learning benchmarking, sensitivity analysis, Monte Carlo experimentation, and reproducible climate data science workflows.

Although the dataset is not intended for operational forecasting or policy applications, it provides a statistically coherent and physically consistent synthetic climatological benchmark for methodological research and teaching. Future developments of the CDSim framework may incorporate additional climatic variables (e.g., humidity, solar radiation), higher temporal resolution, and more advanced multivariate dependence structures to further improve the representation of complex hydro-climatic processes.

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: Davor Kržišnik, University of Ljubljana, Slovenia

Reviewed by: Josias Da Silva Cruz, Federal University of Pará, Brazil