A Hybrid LSTM–LLM Framework for Electricity Load Forecasting with Adaptive Optimization

Abstract

Accurate electricity load forecasting is essential for the reliable operation of modern energy systems, particularly under increasing variability from renewable integration and environmental factors. This paper proposes a hybrid forecasting framework that combines Long Short-Term Memory (LSTM) networks with a GPT-assisted adaptive optimization strategy for short-term load prediction. An initial LSTM model is constructed to capture temporal dependencies in multivariate time-series data, including temperature and humidity. A feedback-driven refinement process is then introduced, in which GPT is utilized as a heuristic optimization component to iteratively adjust model configurations based on observed training behavior and prediction errors. A simulation-based case study is conducted using a dataset designed to reflect realistic load patterns. The results demonstrate consistent improvements across multiple evaluation metrics and show that the proposed method outperforms baseline and benchmark models. The findings suggest that integrating adaptive optimization mechanisms can enhance forecasting accuracy and robustness, while highlighting the potential of large language models as auxiliary components in time-series prediction tasks.