Improved Convolutional Neural Network and its Application in Non-Periodical Runoff Prediction

Due to the influence of human regulation and storage factors, the runoff series monitored at the hydro-power stations often show the characteristics of non-periodicity which increases the difficulty of forecasting. The prediction model based on the neural network can avoid the interference of the non-periodicity by focusing on the relationship between rainfall input and runoff output. However, the physical correlation of the rainfall-runoff and the complexity of the neural network still flaw the subdivision research. In this paper, an improved convolutional neural network (CNN) was innovatively constructed to model runoff prediction, which contains effective layers design and adaptive activation function. The long-term and irregular observation data collected by the Zhexi reservoir were used for training and validation. In addition, the models based on traditional artificial neural networks and ordinary CNN were applied to the forecast simulation for contrast. Evaluation results using real data indicated that the improved CNN model performs better in these acyclic series, with over 0.9 correlation coefficient values and under 185 root means square error values during the validation, meanwhile averting the gradient vanishing and negative discharge problems occurring in other models. Numerous indicators and plots prove the excellent effect and reliability of the model forecast. Considering the robustness and validity of the neural network, this research and verification are of significance to non-periodic reservoir inflow prediction.