Electrical Load Demand Forecasting using LSTM- Kolmogorov Arnold Network model for the Western Grid of India

Accurate short-term load forecasting is essential for maintaining the secure and economical operation of power systems. In large interconnected networks such as India's Western Regional Load Dispatch Centre (WRLDC), variations in electricity demand create challenges for generation scheduling, reserve allocation, and operational planning. Developing forecasting models that can effectively learn both temporal patterns and nonlinear behavior from historical load data remains an active area of research.This work presents a model, a fusion of Long Short-Term Memory network and Kolmogorov Arnold Network (LSTM-KAN), for hourly load forecasting. The proposed approach uses the previous 24 hours of load demand data to forecast the load profile for the next hour using a sliding window forecasting strategy. The LSTM layer is employed to extract temporal information from sequential observations, whereas the KAN component improves nonlinear function approximation through adaptive learnable mappings.The hourly load profiles of WRLDC between the dates 31st December 2020 to 31st December 2024 are used in the model development, while data between 1st January to 30th June 2025 are utilized for testing. The presented model has Mean Absolute Error = 977 MW, Mean Absolute Percentage Error = 1.53% and R² = 0.963 for the test set. It can be observed that the LSTM-KAN model has the ability to capture the load dynamics and give accurate forecasts with reasonable values of the evaluation measures. Since the LSTM-KAN model provides precise predictions and operates efficiently, it can be effectively used in decision-making in power systems.