A novel hybrid GWO-Bi-LSTM-based metaheuristic framework for short-term solar photovoltaic power forecasting
Abstract
Precise solar photovoltaic (SPV) power output forecasting is crucial for improving grid dependability and optimizing energy management systems. This work develops an innovative deep learning-based model for SPV power forecasting, by considering gray wolf optimization (GWO) and whale optimization algorithm (WOA) to enhance performance using three different time horizons, i.e., minutes, hourly, and daily, for real time data collected from a New Delhi, India location which has composite climate, showing adaptability of model's performance in other locations as well. The suggested approach integrates the advantages of various optimization algorithms to improve the training of a deep learning network, i.e., long short-term memory and bi-directional-long short-term memory, particularly by fine-tuning its hyperparameters to get superior forecasting accuracy GWO and WOA are employed to optimize the hyperparameters i.e., epochs and learning rates of the deep learning model, thereby alleviating problems such as local minima and slow convergence. Historical meteorological data, encompassing irradiance, cloud cover, wind speed, wind direction, temperature, and SPV power, are utilized to train the model, which exhibits enhanced performance for forecasting SPV power relative to conventional forecasting techniques. The optimized proposed method is further compared with the deep learning models. The models optimized with GWO and WOA show significant improvements in forecasting accuracy, with the GWO-bi-directional-long short-term memory model obtaining the lowest root mean square error of 0.0154 and the highest R2 value of 0.9988 for 5-min interval data. The modified variations decreased predicting errors by up to 45% compared to baseline models such as long short-term memory and bi-directional-long short-term memory models. These findings demonstrate that hyperparameter adjustment significantly enhances convergence and forecasting precision. The suggested methodology holds significant promise for high-resolution, real-time SPV forecasting, facilitating applications like power dispatch, demand-side management, and the integration of renewable energy into smart grids. This method shows potential for enhancing smart grid technologies and integrating renewable energy.