A stacking ensemble machine learning model for improving monthly runoff prediction

Accurate runoff prediction is significant for many tasks, such as water resource allocation, flood control, disaster reduction, and water conservancy project operation and scheduling in river basins. However, it is difficult to accurately predict using a single model due to the characteristics of runoff series, such as non-stationary, strong randomness, and complexity. We propose a new runoff prediction model based on the Stacking (CNN-BiLSTM) ensemble learning algorithm. This model integrates the echo state network (ESN) model, stacked autoencoder (SAE) model, support vector machines (SVM) model, and long short-term memory (LSTM) model as the base-learner and the CNN-BiLSTM hybrid model as the meta-learner, further improving prediction accuracy through cross-validation. First of all, considering the difference in data observation and training principles of different prediction models, the Spearman correlation coefficient (SCC) method is used to analyze the correlation between the prediction errors of multiple single models and the mean absolute percentage error (MAPE) is used to evaluate their prediction ability. The model with low correlation and strong prediction ability is selected as the first layer prediction model of the stacking ensemble algorithm, that is, the base-learner. Then, the prediction results of the base-learner are used as a new input feature set and passed on to the second layer prediction model of the stacking ensemble learning model, namely the meta-learner, for training and prediction to obtain the final monthly runoff prediction value. To verify the prediction accuracy and generalization ability of the Stacking (CNN-BiLSTM) ensemble model, Manwan hydropower station in southwest China, Xiajiang hydrological station in eastern China, and Wangjiahui hydrological station in northern China are selected as research objects. The model’s performance is assessed using four evaluation metrics and compared against eight different models, and the prediction effect at the peak is tested using the peak evaluation indicator (EP). The experimental findings indicate the RMSE of the proposed Stacking (CNN-BiLSTM) model at three stations is reduced by 37.8%, 78.9%, and 53.6%, respectively, compared with the single model LSTM in the test phase; NSE increased by 11.45%, 112.41%, and 80.55%, respectively. Among the 15 peak predictions, the Stacking (CNN-BiLSTM) model shows the best EP on 12 peaks, demonstrating that the model’s peak prediction results are closest to the actual peaks. The Stacking (CNN-BiLSTM) proposed in this article provides a new method for predicting monthly runoff and peak values.