High-Precision Prediction of Hydrological Time Series Based on a MultiModal-TCN-LSTM Fusion Architecture

In response to the problems that traditional hydrological time series prediction methods are unable to capture nonlinear features, long-term dependencies, and fail to fully utilize the collaborative information of multiple elements, this paper proposes a high-precision prediction model based on the MultiModal-TCN-LSTM (Multi-modal Time Series Convolutional - Long Short-Term Memory Network). This model innovatively constructs a dual-branch parallel architecture: extracting multi-scale local features of water level and flow data through the Time Series Convolution Network (TCN), modeling the entire long-term dependencies using the Long Short-Term Memory Network (LSTM), and mining the deep correlations among hydrological elements through the multi-modal fusion layer. The data is preprocessed using Z-score standardization and linear interpolation methods. Based on the 5-year measured water level-flow data from the Batang Hydrological Station in the upper reaches of the Jinsha River, the optimal historical input length of 8 hours (h) was determined through time window optimization. In the prediction experiments of 12h, 24h, and 48h ahead, the comprehensive evaluation using RMSE, MAE, MAPE, and R² indicates that this model significantly outperforms the baseline models such as LSTM, TCN, TCN-LSTM, and MultiModal-LSTM, verifying the effectiveness of the multi-modal fusion and TCN-LSTM hybrid architecture. Especially in the long-term prediction period, the TCN component significantly alleviates the performance degradation of the model through the expansion convolution mechanism. This research provides a new idea for hydrological prediction and has important application value for hydropower station scheduling, water resource allocation, and flood control and disaster reduction.