Dynamic event-aware load forecasting: a hybrid framework integrating a RAG-agent with time series models

Short-Term Load Forecasting (STLF) is crucial for maintaining grid stability. However, traditional methods that rely mainly on historical patterns often fail to capture the irregular fluctuations in real-world events. To address this challenge, this paper proposes Hy-RAG-Forecast, a hybrid forecasting framework that tightly integrates deep learning–based load prediction with adaptive large language model (LLM)–driven event inference. In the proposed architecture, a Long Short-Term Memory (LSTM) network serves as the baseline predictor to capture regular temporal dependencies in historical load data. Upon the occurrence of potential abnormal conditions, an LLM based event analysis agent is triggered to augment the forecasting process. The agent is implemented using a Retrieval-Augmented Generation (RAG) paradigm within the LangChain framework, enabling it to retrieve real-time, domain-relevant information from external data sources, such as online event listings and public activity announcements. It discovers outdoor activities such as concerts or sports competitions. Then, it estimates how much impact these events may have on demand. As a result, it has created new features that update over time to help make predictions more accurate. This method was tested on the UK public electricity demand dataset. The results indicate that Hy-RAG-Forecast works better than ARIMA, a regular LSTM, and an LSTM that uses fixed text features. A case study shows the system can respond to unusual changes caused by events. The agent also explains how it updates the forecast, so the changes are clear to users. Overall, Hy-RAG-Forecast combines quantitative historical data with real-time contextual intelligence, offering a more adaptive and explainable approach to shortterm load forecasting.