Abstract: Aiming at the problems of long cycle and high cost of water quality detection, a RF-Softmax water quality prediction model based on genetic algorithm optimization was proposed. Using machine learning method and random forest algorithm for feature selection, the mathematical relationship model between ammonia nitrogen and total phosphorus in water samples and water quality categories was established to predict water quality categories. The genetic algorithm was used to replace the gradient descent method used in the training process of the traditional Softmax regression algorithm, which solved the problem that the logistic regression algorithm was easy to fall into the local optimal solution when the objective function was not strictly convex. The surface water in Nanzha street of Jiangyin city was used as the research object for verification. The results showed that the RF-Softmax regression model optimized by GA had the highest prediction accuracy. Compared with the traditional Softmax regression and BP neural network, the prediction accuracy increased by 11.73% and 8.40%, respectively, the mean error decreased by 58.68% and 34.92%, and the average root mean square errors decreased by 39.02% and 23.62%, respectively. The optimization effect is remarkable, which can realize efficient, accurate, low cost and fast surface water quality prediction. It provides a new idea for water quality monitoring and early warning, and is of great significance for water quality management and environmental protection.

Key words: water quality prediction, prediction model, genetic algorithm, Softmax regression, random forest, machine learning