Large Language Model-Empowered Interactive Load Forecasting

TL;DR

The paper addresses the need for human-in-the-loop load forecasting by introducing an LLM-powered multi-agent framework that orchestrates the entire forecasting pipeline. It combines a Task Manager, Preparation Assistant, Model Manager, Model Developer, and Deployment Operator to enable natural-language interaction and stage-wise control, guided by an interactive Bayesian optimization loop. The approach yields improved forecasting accuracy when users contribute insights and shows feasible token-based costs for real-world deployment. This framework has practical implications for power-system operators, offering an adaptable, interpretable, and cost-conscious path to integrating advanced forecasting methods with domain expertise. Future work aims to enhance memory efficiency and provide explainable AI capabilities to further boost trust and adoption.

Abstract

The growing complexity of power systems has made accurate load forecasting more important than ever. An increasing number of advanced load forecasting methods have been developed. However, the static design of current methods offers no mechanism for human-model interaction. As the primary users of forecasting models, system operators often find it difficult to understand and apply these advanced models, which typically requires expertise in artificial intelligence (AI). This also prevents them from incorporating their experience and real-world contextual understanding into the forecasting process. Recent breakthroughs in large language models (LLMs) offer a new opportunity to address this issue. By leveraging their natural language understanding and reasoning capabilities, we propose an LLM-based multi-agent collaboration framework to bridge the gap between human operators and forecasting models. A set of specialized agents is designed to perform different tasks in the forecasting workflow and collaborate via a dedicated communication mechanism. This framework embeds interactive mechanisms throughout the load forecasting pipeline, reducing the technical threshold for non-expert users and enabling the integration of human experience. Our experiments demonstrate that the interactive load forecasting accuracy can be significantly improved when users provide proper insight in key stages. Our cost analysis shows that the framework remains affordable, making it practical for real-world deployment.

Figures (8)

• Figure 1: The proposed multi-agent collaboration framework includes a Task Manager that handles user interaction and workflow control and other specialized agents manage different tasks in the pipeline.
• Figure 2: Communication mechanism of the multi-agent framework, illustrating the topic-based messaging topology and a simplified example of the Model Manager sending a message to the Model Developer.
• Figure 3: A simplified system prompt of the Task Manager with four core components: profile, memory, workflow, and actions.
• Figure 4: Workflow of the task preparation stage. User interactions and system checks are conducted at each step to ensure correct task definition and data readiness.
• Figure 5: Example workflow of the model training and evaluation stage. The user's input is first filtered and processed by the Task Manager, then forwarded to the Model Manager. The Model Manager follows a two-stage reasoning and action process before passing the resulting decision to the Model Developer, who is responsible for executing the corresponding function calls to train and evaluate the model. The evaluation results are then visualized and presented to the user.