Optimization of LSTM Gas Concentration Prediction Based on Improved Multi-Strategy Beluga Algorithm

For the purpose of thoroughly exploiting the information provided by coal mine safety monitoring and achieving better accuracy in gas concentration prediction, a multistrategy modified Beluga Algorithm (MBWO) was proposed to optimize the short-duration memory recurrent neural network (LSTM) gas concentration prediction model. Firstly, incorporating good point set initialization, an elite pool strategy, dynamic reverse learning, and golden sinusoidal strategy enhances the algorithm. Its effectiveness and feasibility are demonstrated through testing with different test functions. Secondly, to develop a gas concentration prediction model, the pre-processed time series data of gas concentration and related factors were used as input for the LSTM model, optimized by the improved Beluga algorithm. Experimental data were used as samples to compare the proposed model with BP, LSTM, PSO-LSTM, and BWO-LSTM models. The comparison showed that the proposed model exhibited better prediction performance, with an average absolute percentage error of 0.7934%, a root mean square error of 0.0076, and a coefficient of determination of 0.9985. This indicates that the proposed gas concentration prediction model provides higher prediction accuracy and significantly enhances prediction accuracy.