Water Quality Predict Based On Machine Learning Technology

Water quality prediction is a basic work in water resource management and pollution control. It is crucial to predict water quality accurately. Currently there are a variety of domestic and international water quality prediction methods, but these methods with some drawbacks. Four kinds of water quality prediction models are discussed in this thesis, they are the water quality parameter time series forecasting model with support vector regression, relevance vector machine regression, online sequential extreme learning machine and deep belief network.Biogeography-based optimization and its improved method are proposed for support vector machine prediction model. And the model is applied to predict the four important water quality parameters, which are PH, dissolved oxygen, permanganate index and ammonia nitrogen. Panzhihua Cave automatic water quality monitoring time series data released by the Ministry of Environmental is used to validate the effectiveness of the new control variables optimization method and then compared with the traditional control variables optimization methods, experimental results indicate that the improved biogeography-based optimization can give a better prediction result.There are some disadvantages on support vector machine water quality prediction model, such as ’Mercer’ kernels, the number of support vectors will increase linearly with the increase of training samples, and no probability output. On this basis, water quality parameter time series forecasting model based on a relevance vector machine regression is proposed and validated, by comparing with the support vector machine regression forecasting model. The linear function and gaussian function are selected as the kernel function to contrast the result of different kernel functions, and gives the confidence level of 95% confidence intervals in the application of the relevance vector machine regression forecasting model. Experimental results indicate that the relevance vector machine regression forecasting model is no less than the support vector machine regression forecasting model. Furthermore, the relevance vector machine regression can give prediction values and also compute the confidence levels for prediction results.Artificial neural networks algorithms prone to over-or under-study learning, local minima, difficult to determine the network structure, and poor generalization ability. According to the characteristics of online monitoring of water quality parameters, a new algorithm based on an online sequential extreme learning machine time series water quality parameter forecasting model is proposed in this thesis. The same experimental data as in support vector regression model is used to validate the validity of the model, and then compared with the artificial neural network prediction model. Experimental results show that the online sequential extreme learning machine is superior to the artificial neural networks with better prediction accuracy and shorter training time.Besides, water quality parameter time series forecasting model based on a deep belief network is discussed.