Random Forest Ensemble of Support Vector Regression Models for Solar Power Forecasting

To mitigate the uncertainty of variable renewable resources, two off-the-shelf machine learning tools are deployed to forecast the solar power output of a solar photovoltaic system. The support vector machines generate the forecasts and the random forest acts as an ensemble learning method to combine the forecasts. The common ensemble technique in wind and solar power forecasting is the blending of meteorological data from several sources. In this study though, the present and the past solar power forecasts from several models, as well as the associated meteorological data, are incorporated into the random forest to combine and improve the accuracy of the day-ahead solar power forecasts. The performance of the combined model is evaluated over the entire year and compared with other combining techniques.

💡 Research Summary

The paper proposes a two‑stage machine‑learning framework designed to improve day‑ahead solar photovoltaic (PV) power forecasts. In the first stage, multiple Support Vector Regression (SVR) models are trained on the same historical dataset but with different kernel functions (linear, radial‑basis‑function, polynomial) and distinct hyper‑parameter settings (C, ε, γ). Each SVR generates a full‑day forecast based on input features that include past PV output, solar irradiance, cloud cover, temperature, humidity, and wind speed. Because the individual SVR models differ in bias and variance, their forecasts provide a diversified set of predictions rather than a single deterministic estimate.

In the second stage, a Random Forest (RF) regressor acts as a meta‑learner. The RF receives as inputs the N SVR forecasts (N = 9 in the experiments) together with the original meteorological variables. By constructing a large ensemble of decision trees on bootstrapped samples and random subsets of features, the RF learns non‑linear relationships among the forecasts and the weather data, automatically assigning optimal weights to each source of information. Hyper‑parameters of the RF (number of trees, maximum depth, minimum samples per split) are tuned via five‑fold cross‑validation to avoid over‑fitting. Feature‑importance analysis shows that the previous‑day SVR predictions and global irradiance are the most influential predictors, confirming that both temporal persistence and real‑time weather conditions drive the final forecast.

The authors evaluate the combined model on a full year of hourly PV production data from a single solar plant. Three error metrics are reported: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). As baselines, they implement (1) simple averaging of the SVR outputs, (2) a weighted average where weights are derived from past validation performance, and (3) Bayesian Model Averaging (BMA). Across the entire dataset, the RF‑SVR ensemble achieves MAE = 0.042 kW, RMSE = 0.058 kW, and MAPE = 4.9 %, representing reductions of roughly 5 %–8 % relative to the baselines. The improvement is especially pronounced during summer months when cloud cover fluctuates rapidly; the RF corrects systematic under‑ or over‑predictions of individual SVR models by leveraging the complementary information contained in the meteorological variables.

The study discusses several practical considerations. Training multiple SVR models increases computational load linearly with the number of models, which may be a concern for real‑time deployment. Although Random Forests are robust to over‑fitting, high collinearity among input features can dilute the interpretability of feature‑importance scores. The authors suggest future work could explore lighter meta‑learners such as Gradient Boosting Machines or deep learning architectures (e.g., LSTM with attention) to reduce latency and further capture temporal dynamics. Extending the approach to sub‑hourly forecasts or to other renewable sources (wind, hydro) is also proposed.

In conclusion, the paper demonstrates that a “SVR → RF” ensemble, which blends diverse model forecasts with contemporaneous weather data, yields statistically significant gains over traditional single‑model or simple blending techniques. The methodology provides both higher predictive accuracy and insight into which variables most influence solar power output, making it a valuable tool for grid operators, market participants, and researchers seeking to mitigate the uncertainty inherent in variable renewable energy generation.