Energy Production Forecasting based on Stacking Model and SHAP

With the growth of global energy demand and the wide application of renewable energy, accurate prediction of energy production can help market participants and decision makers to optimise resource allocation and improve energy efficiency. However, since energy production is affected by a variety of environmental variables, in order to improve the prediction accuracy and stability, this study adopts three machine learning methods, namely, Random Forest, CatBoost, and Stacking Model, to make predictions, and analyses their performances through experiments. The study first preprocesses the dataset, including data cleaning, normalisation and feature engineering, and adopts the strategy of 80% training set + 20% testing set for model training. Subsequently, evaluation metrics were calculated separately, and interpretability analysis was performed using SHAP to reveal the effects of key variables on biogas production. The experimental results showed that the stacked model performed the best on all evaluation metrics, with R² = 0.81821 and RMSE = 0.07253, which were significantly improved compared to CatBoost (R² = 0.69796, RMSE = 0.0935) and Random Forest (R² = 0.69682, RMSE = 0.55009). In addition, the results of SHAP analyses further showed that temperature was the most critical factor affecting biogas production and significantly interacted with humidity and external temperature. The experimental results demonstrated the efficiency and stability of the stacked model and verified its potential application in biogas yield prediction.