A CSO-based VAE-LSTM Model in Hourly Solar and Load Forecasting for Enhancing the PV-BES Performance Considering Battery Degradation Cost

This study presents a two-stage battery degradation optimization framework incorporating solar PV and load forecasting. In the first stage, a new deep learning model that uses a Variational Autoencoder (VAE) and Long Short-Term Memory (LSTM), improved by Competitive Swarm Optimization (CSO), is suggested for predicting data on an hourly basis. VAE decomposes input data, while CSO optimizes LSTM hyperparameters. The model, tested on two input sets with cloud cover considerations, demonstrates superior performance based on RMSE, MAE, and R² metrics. The results of the proposed model are compared with some conventional models, such as SVR, XGBoost, and LSTM, tested by the author, as well as with some existing state-of-the-art literature. As compared to the tested models, the proposed model with Type B feature set shows an average reduction in RMSE of 55.46% and an average improvement in accuracy of 6.08%. When the proposed model is compared with existing literature, the average improvement in accuracy is 3.386%. In the second stage, battery degradation cost is minimized by determining the optimal battery capacity and scheduling the battery energy storage system's State of Charge (SoC) using CSO. Under various weather and SoC conditions, CSO identifies an optimal capacity of 29.742 kWh and an initial SoC of 50–60%. Optimal degradation costs are achieved across different weather conditions, validating CSO's superiority.