MULTI-SITE SOLAR ENERGY FORECASTING USING XGBOOST WITH ADVANCED FEATURE ENGINEERING AND SHAP EXPLAINABLE AI

Abstract

The accurate forecasting of solar energy is of considerable importance for the successful integration of solar power into the power grid and effective power grid management. However, temporal variations, uncertainty of weather and geographical diversity are major challenges for generating generalized forecast-ing models, as these models are required to be intermittent. To this end, this paper introduces explainable ML Methods for multi-site solar energy forecasting that integrate powerful temporal feature engineering techniques and SHapley Additive exPlanations (SHAP) with Extreme Gradient Boosting (XGBoost). The proposed framework utilizes historical PV generation data and historical meteorological data from four locations across Australia with different geographic characteristics: Mt. Gambier (South Australia), Eddystone Point (Tasmania), Mildura (Victoria), and Albany (Western Australia). In this study, PV energy is forecasted one hour-ahead: the PV output in the past, including the PV generation data and meteorological observation, are used to predict the PV output at the next time. A complex preprocessing process was designed, such as missing values imputation, data normalization, temporal feature extraction, lag-based feature generation, rolling statistical analysis, site encoding, feature interaction construction, etc. All the datasets were divided in a chronological split strategy where 70%, 15%, and 15% of the data were put aside for training, validation, and testing. The experimental results demonstrate the effectiveness of the proposed framework with XGBoost model in predicting PV energy output compared to other models in different geographic areas. Further, the knowledge gained from the interpretable analysis provided by SHAP makes the model more transparent and increases the confidence of the operator when assessing the impact of the meteorological and temporal variables on the model to predict solar power generation. The proposed explainable forecasting framework is an effective solution for intelligent management of renewable energy and trustworthy decision support of smart grid.