ELECTRICITY DEMAND FORECASTING FOR POWER GENERATION ACROSS MULTIPLE REGIONS IN INDIA

Forecasting electricity demand is an important factor in maintaining grid stability, optimizing generation resources, and sustainable energy planning, especially in large-scale and heterogeneous power systems like India. This paper introduces a powerful, AI-based electricity demand forecasting system that incorporates multi-source data and state-of-the-art deep learning models to overcome the shortcomings of conventional statistical and standalone machine learning models. The suggested system uses real-time NASA POWER meteorological data and past demand data provided by Ember Energy and then undergoes extensive preprocessing, time synchronization, and feature engineering, such as lag variables, rolling statistics, and cyclical time encoding. A hybrid ensemble model of Neural Hierarchical Interpolation of Time Series (NHiTS) and iTransformer is created to be able to capture both time dynamics and multivariate dependencies. The NHiTS component is used to model multi-scale time dynamics (e.g., hourly, daily, and seasonal variations) and the iTransformer is used to model complex cross-variable relationships between weather parameters and electricity demand. A learnable fusion layer combines the outputs of both models to boost predictive accuracy. Experimental analysis shows that the proposed ensemble model has a Mean Absolute Percentage Error (MAPE) of 11.4% and an average forecasting error of 89.2% that is better than the single models and the traditional benchmarks. The coefficient of determination is also high (R 2>0.88) which means that the system has a high level of agreement between the predicted and actual demand. Moreover, the locality-aware scaling enhances the regional forecasting up to 12 times, which demonstrates the significance of spatial heterogeneity in demand modeling. Besides prediction, the system also has an intelligent energy recommendation module, which estimates the best distribution of energy mixes based on the predicted demand and weather conditions, with a recommendation accuracy of 92.5. The framework is implemented with the help of an interactive dashboard, which allows visualizing in real-time, multi-horizon forecasting (up to 30 days), and providing grid operators with decision support. On the whole, the suggested system proves to be very accurate, scalable, and practically applicable to the management of smart grids, which will lead to the increase in the efficiency of operations, decrease in the wastage of energy, and the further integration of renewable energy sources.