A Comparative Analysis of ANN, LSTM and Hybrid PSO-LSTM Algorithms for Groundwater Level Prediction

The accurate prediction of groundwater levels is vital for effective water management strategies due to its significance in various critical sectors such as drinking water supply, agriculture, ecosystem support, industrial applications and urbanization. The objective of research is to develop a robust groundwater levels forecasting model using traditional artificial neural network (ANN), standalone long short-term memory (LSTM), and integrating an optimization approach based on particle swarm optimization (PSO). Initially, ANN and a standalone LSTM model was evaluated on a dataset spanning 29 years from the Chhattisgarh state in India. The LSTM model yielded moderate accuracy with statistical metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), R-squared (R2) and Cosine Similarity were 1.0924, 1.4565, 1.2068, 0.853, 0.924 respectively. Subsequently, a Hybrid PSO-LSTM Algorithm was proposed and implemented to enhance the accuracy of groundwater level predictions. This hybrid approach combines the strengths of PSO optimization with LSTM modeling. PSO optimization algorithm is used to get the optimized value of various hyperparameters that affects LSTM. Evaluation of the hybrid PSO-LSTM model demonstrated significant improvement over the standalone LSTM, achieving an MAE, MSE, RMSE, R2 and Cosine Similarity obtained as 0.182, 0.062, 0.249, 0.947, 0.989 respectively. These results show that the implemented hybrid PSO-LSTM algorithm outperforms both traditional ANN and standard LSTM. The accurate prediction of groundwater level leads to effective management and planning for various purposes such as water harvesting, agriculture, industrialization etc. Therefore, these findings can significantly contribute to predicting the groundwater level effectively.