Very Short-Term Load Forecasting Using Artificial Neural Networks with Meteorological Variables: A Case Study of Uttarakhand

Abstract

This study presents an Artificial Neural Networks (ANN) based Very Short-Term Load Forecasting (VSTLF) framework for Uttarakhand using Artificial Neural Networks (ANN) with the inclusion of meteorological variables. The dataset comprised hourly load demand from the State Load Dispatch Centre (SLDC) and meteorological parameters (temperature, humidity, and dew point) obtained from NASA POWER LAARC for the period January 2020 to April 2024. Multiple ANN configurations were evaluated with varying lag inputs, hidden neurons, and training algorithms. The results indicate that incorporating meteorological variables significantly enhances forecasting accuracy, with the best-performing model (16-hour lag, 96 hidden neurons, Bayesian Regularization) achieving a test RMSE of 60 MW, MAPE of 2.66%, and R² of 0.95. Residual, histogram, and scatter analysis confirmed unbiased predictions with strong agreement between actual and forecasted loads. The findings demonstrate the critical role of weather-sensitive inputs in improving load forecasting in climatically diverse regions like Uttarakhand, where temperature and humidity patterns strongly influence electricity demand. This study advances reliable short-horizon forecasting for grid stability, renewable integration, and operational planning in emerging power systems.