Ensemble Machine Learning Technique Based on Gaussian Algorithm for Stream Flow Modelling

Abstract

Streamflow modelling is regarded as a crucial part of managing and planning water resources. Water resources engineers face a variety of challenges when predicting streamflow. These difficulties are caused by complex natural processes that involve non-linearity, non-stationarity, and randomness. This research investigates the application of machine learning (ML)-based models for forecasting streamflow discharge (Q) using input variables, including temperature and rainfall. The study attains the essential stationarity required for precise modelling using Augmented Dickey-Fuller tests, data normalization, and transformation. In contrast, unit root tests identify initial-level non-stationarity and call for first-differencing. Correlation matrix analysis identifies relevant input combinations. The result findings are supported by statistical metrics including mean squared error (MSE), mean absolute error (MAE), and Pearson correlation coefficient (PCC). Notably, in terms of prediction accuracy, the Gaussian Process Regression (GPR) GPR-M3 model stands out as a notable performer, with a low MAE value of 0.034 in the calibration phase and 0.027 in the verification phase. The success of these techniques is further supported by first and second-order ensemble algorithms, with some models reaching a perfect PCC score during both the calibration and verification phases. The study emphasizes the significance of preprocessing, model selection, and ensemble procedures in improving the accuracy of streamflow prediction models. In addressing complex nonlinear interactions, artificial intelligence (AI)- based models are valuable tools for both technical applications and practical understanding.