Optimized neural network to predict the experimental minimum period of coal spontaneous combustion

The harmful gases produced from coal spontaneous combustion (CSC) can cause the environmental pollution. Being able to predict the experimental minimum period of CSC (EMPCSC) is essential in controlling CSC and effectively reducing harmful gas emissions. To obtain high prediction accuracy, we used three optimization algorithms, namely the genetic algorithm (GA), ant colony algorithm (ACO), and particle swarm optimization algorithm (PSO), to optimize the backpropagation neural network (BPNN). R², MSE, RMSE, and MAPE were used as evaluation indexes to determine the most accurate prediction model for EMPCSC. Data of 424 coal samples from 15 regions in China were analyzed, with 207 and 217 samples having a spontaneous combustion period of less than 40 days (W) and more than 40 days (V), respectively. The two groups were further distributed between low-temperature slow oxidation (W₀ and V₀) and low-temperature fast oxidation (W₁ and V₁). The results indicated that the prediction performance of the BPNN model optimized using PSO (PSO-BPNN) was better than that of the GA-BPNN and ACO-BPNN models. After optimization through PSO, the goodness of fit (R²) of groups W₀, W₁, V₀, and V₁ increased from 0.9180, 0.8746, 0.9987, and 0.9782 to 0.9857, 0.9639, 0.9997, and 0.9994, respectively. Therefore, the results can provide a theoretical reference for selecting the optimal neural network model to predict EMPCSC with high accuracy.