A novel TCN-augmented CNN-LSTM architecture for accurate monthly runoff forecasting
摘要
To address the randomness and non-linear aspects of runoff data, this research introduces an efficient runoff forecasting model aimed at improving the accuracy of monthly runoff forecasts and addressing the challenges posed by non-linear patterns and long-term dependencies. Considering the constraints of conventional runoff forecasting techniques in identifying extended temporal correlations and regional trends, this research innovatively introduces the TCN-CNN-LSTM (TCL) runoff prediction model, which integrates Temporal Convolutional Network (TCN) layers, Convolutional Neural Network (CNN) layers, and Long Short-Term Memory (LSTM) networks. This model effectively addresses the challenges of local feature extraction and the gradient vanishing problem in long sequences of time series data. First, the TCN layer is utilized to discern extended temporal correlations through dilated causal convolutions to tackle the non-linearity and persistent interdependencies within sequential data forecasting. Next, the CNN layer, with kernels of varying sizes, is used to capture local features at different scales. Finally, the LSTM layer employs a specialized gating mechanism to manage long-term dependencies within the sequential information. To verify the effectiveness of the TCL model, we conducted tests on datasets from three hydrological stations. The results demonstrate that the TCL model excels in predictive performance and is capable of effectively capturing local trends and long-term dependencies in sequential data. Specifically, at the Jiayuguan hydrological station, the TCL model achieved a reduction of 148.33%, 93.20%, and 6.22% in MAE, RMSE, and MAPE, respectively, compared to the TCN-CNN model. At the Hongshan River hydrological station, the TCL model achieved a determination coefficient (