Improved runoff forecasting based on time-varying model averaging method and deep learning

Abstract

In order to improve the accuracy and stability of runoff prediction. This study proposed a dynamic model averaging method with Time-varying weight (TV-DMA). Using this method, an integrated prediction model framework for runoff prediction was constructed. The framework determines the main variables suitable for runoff prediction through correlation analysis, and uses TV-DMA and deep learning algorithm to construct an integrated prediction model for runoff. The results demonstrate that the current monthly runoff, the runoff of the previous month, the current monthly temperature, the temperature of the previous month and the current monthly rainfall were the variables suitable for runoff prediction. The results of runoff prediction show that the TV-DMA model has the highest prediction accuracy (with 0.97 Nash-efficiency coefficient (NSE)) and low uncertainty. The interval band of uncertainty was 33.3%-65.5% lower than single model. And the prediction performance of the single model and TV-DMA model in flood season is obviously lower than that in non-flood season. In addition, this study indicate that the current monthly runoff, rainfall and temperature are the important factor affecting the runoff prediction, which should be paid special attention in the runoff prediction.

Integrated prediction of water pollution and risk assessment of water system connectivity based on dynamic model average and model selection criteria

“…With the rapid development of artificial intelligence and data science, a series of advanced machine learning methods have been successfully applied to research on water pollution prediction. However, the applicability of different methods may vary greatly [ 23 ]. Therefore, this study selected candidate models suitable for water pollution prediction from the perspective of method applicability and used the selected candidate models to construct water pollution prediction models.…”

“…This change may cause significant loss of local prediction accuracy [ 19 , 20 ]. In addition, some research found that if the prediction results of all candidate models were overestimated or underestimated, the performance of the model average method would be significantly lower than that of the model selection method [ 16 , 23 ]. The main reason is that the model average is a weighted average of the candidate models.…”

Application of the Improved K-Nearest Neighbor-Based Multi-Model Ensemble Method for Runoff Prediction

“…When training data feature attributes are chosen collectively, the predicted model data points at different time points may not align with the considered collective attributes [30]. Consequently, relying solely on the performance of a single model during the training phase to determine corresponding weights may lead to significant errors in localized predictions [31]. This approach disregards the influence of the quantity and quality of training period data features on the effectiveness of the ensemble forecasting process during the weight determination phase.…”