Optimizing LSTM Hyperparameters with Fish Swarm Optimization for Enhanced Power Load Prediction in Wireless Networks

Abstract

Predicting power loads is an important part of managing the electricity system and provides a basic guarantee for the dependability and financial operation of state grid companies. In order to minimise energy production costs, it is advantageous to have an accurate prediction for efficient energy scheduling, which involves balancing power generation and demand. Many scholars have devoted their time and energy to creating trustworthy load forecasting models in the hopes of achieving the highest possible prediction accuracy. Utilising a variety of ML, DL, AI, and hybrid approaches allows for accurate power load prediction. An integrated model for power load prediction using the Fish Swarm Optimisation (FSO) algorithm and the Long Short-Term Memory (LSTM) neural network is presented in this research. By optimising the LSTM network's hyper parameters using FSO, this study speeds up model convergence and avoids becoming stuck in local optima, which reduces the effect of humans picking LSTM hyper parameters at random on prediction results. The proposed approach is evaluated using mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE) and R2 metrics and compared with LSTM, CNN, Attention-LSTM, AC-BiLSTM and LSTM-PSO approaches. The proposed LSTM-FSO significantly reduces MAE between 3.75% - 49.59%, MAPE between 9.98% - 39.65%, RMSE between 11.17% - 56.10% and increased R2 between 0.1% - 8.07% compared to other methods. The results of the experiments show that the suggested system has great potential for use and significantly beats competing solutions. Power load forecasting is shown to be more stable and reliable using the FSO algorithm with LSTM, according to the simulation findings.