Water Quality Prediction Based on CEEMDAN-LSTM-LINREG Model

In this paper, CEEMDAN is introduced to perform modal decomposition of indicator data to capture the characteristics of the data at different time scales. Each feature is constructed as a stacked LSTM sub-prediction model. Since different features represent different dimensions, the cumulative sum of the results of each sub-model as the final prediction result will lead to the problem of error superposition. In this paper, a linear regression model is introduced to self-learn the corresponding weights of each sub-model, and the weighted sum method is used to fuse the results of each sub-model to alleviate the error superposition problem. The dissolved oxygen, ammonia nitrogen, PH and permanganate indicators are taken as examples to verify the prediction model proposed in this paper. The experimental results show that in the models constructed by CEEMDAN and EMD decomposition algorithms, the prediction accuracy of each index is improved by 11.88%, 4.38%, 1.33%, 3.38 and 19.29%, 3.65%, 0.61 %, 2.65% respectively after introducing the linear regression model. It shows that the combined model constructed in this paper has strong generalization ability and robustness.