Integrating Advanced Feature Extraction with Deep Learning Models for Accurate Forecasting of Peak Load Demand and Solar Power Generation

The efficient forecasting of electricity demand and solar power generation is crucial for addressing India's growing energy needs and advancing its renewable energy goals. This study focuses on peak load demand and solar power generation forecasting for the Indian states of Gujarat and Rajasthan using daily time series data collected from 2019 to 2024. The study leverages two advanced feature extraction techniques, seasonaltrend decomposition using loess (STL) and discrete wavelet transform (DWT), to enhance the performance of deep learning models by isolating trends, seasonality, and high-frequency variations. In this work, deep learning models such as one-dimensional convolutional neural network (1D-CNN), long short-term memory (LSTM), gated recurrent unit (GRU), and hybrid architectures were trained and evaluated using root mean squared error, mean absolute error, mean absolute percentage error, and R2 score. The obtained results demonstrate that combining deep learning architectures with feature extraction methods such as STL and a combination of wavelet transform and STL (DWT-STL) improves the forecasting accuracy of standalone deep learning models. Comparative analysis found feature-enhanced CNN and GRU to be the best performing models with lower error metrics and higher R2 scores for both peak load demand and solar power generation forecasting. The proposed approach offers a robust framework for energy forecasting and can guide efficient renewable energy planning to support India's transition to a sustainable energy future.