XAI-Based Photovoltaic Energy Management Framework for Smart Cities

The demand for explainable energy forecasting models has been emphasized in recent studies to enhance decision-making transparency. The efficient conversion of solar energy is stochastic and depends on environmental conditions. To address this, we propose a model-chain framework named XPEMM (XAI-based Photovoltaic Energy Management Model chain) for Solar Energy Production and Management in Smart Cities, ensuring explainability using LIME. Our model chain consists of three stages: Long-term forecasting of Global Horizontal Irradiance (GHI) as the first stage, Regression Modeling of GHI with power as the second stage, and Model-Agnostic explanation generation as the third stage. A recursive forecasting strategy was employed, leveraging Recurrent Neural Networks (RNN) to reduce error accumulation over a 5-year forecast horizon. The output of the proposed model chain assists grid cell operators in real-time monitoring, resulting in optimized performance by predicting the causes of high and low performance in advance, allowing for necessary adjustments. It also aids in site selection for high-budget PV-plant installations. We used two different datasets for training and benchmarking the Recursive Multistep GHI Forecasting Model (RMGFM). The proposed recursive forecasting strategy has been evaluated using RMSE and MAE error metrics which are 91.40 and 47.32 respectively, which are 2.9% 29.39%, and 7.9% less than the previously established approach using LSTM, PCR, and SVR based on RMSE measure and 15.9%, 49.9%, 34.8% less based on MAE metric. This study concludes that a recursive approach is best for long-term forecasting of GHI in regions with cyclic climatic patterns based on RMSE and MAE values. The interpretation of the LIME output establishes the fitness of model agnosticity in the Photovoltaic model chain.