Short-term solar irradiance forecasting using deep learning models and an agent-based decision support framework for localized energy management

Abstract

Accurate forecasting of solar irradiance is critical to supporting the stability of the power grid and allowing for efficient use of renewable energy resources. As solar generation is highly dependent upon weather; therefore, accurate forecasting of short-term solar totals, or Global Horizontal Irradiance (GHI), is useful to energy planners in order to manage power supply and minimize uncertainties around actual operations. This research examines short-term GHI prediction by utilizing high-resolution, ground-based meteorological observations collected from 9 locations throughout Pakistan. The first modelling approach employs three deep learning architectures Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Networks (TCN) to capture temporal dependencies in multi-station weather data. Transformer-based and advanced time series forecasting techniques (i.e. Informer, Autoformer, N-BEATS) have been evaluated to determine how to conduct a more comprehensive evaluation of contemporary solar irradiance forecasting models. The dataset covers the period 2014–2017 with 10-minute resolution, making it suitable for fine-grained short-term forecasting. A comprehensive preprocessing pipeline is applied, including outlier handling, feature engineering, cyclical temporal encoding, and robust scaling. Time-series sequences of 24 time steps are used as input to predict the next-step irradiance value. In addition to centralized multi-station modelling, a second experimental setup explores a multi-agent inspired forecasting framework for single-station prediction. In this configuration, a forecasting agent uses recurrent models (RNN and LSTM) to predict the next-hour irradiance, while a secondary decision agent uses these predictions within a rule-based energy management simulation to evaluate potential battery charging and grid-usage actions. Experimental results show that the optimized stacked LSTM model achieves the best performance for multi-station forecasting, with an $$R^2$$ score of 0.9907 and lower MAE and RMSE compared with GRU and TCN models. For the single-station setup, the agent-based LSTM demonstrates stable forecasting performance under localized weather dynamics. Comparative analysis with recent studies highlights the effectiveness of recurrent deep learning models for short-term solar irradiance forecasting. The proposed framework can assist utility operators and energy planners in improving solar resource management in regions with high solar potential.