A new scheme for probabilistic forecasting with an ensemble model based on CEEMDAN and AM-MCMC and its application in precipitation forecasting

Precipitation affects the generation of runoff and concentration of water resources in basins. The randomness of precipitation contributes to the difficulty and uncertainty of forecasting it. To improve precipitation forecasting accuracy and account for this uncertainty, a new scheme for probabilistic precipitation forecasting is proposed. In the scheme, first, signal decomposition techniques (complete ensemble empirical mode decomposition with adaptive noise) are used to decompose original precipitation series into subsequences. Second, empirical approaches (time series analysis model, grey self-memory model and long-short-term memory) are used to produce a quantitative precipitation forecast. Third, an ensemble model is used to assemble the outputs of empirical approaches, whose weights are determined by the Adaptive Metropolis-Markov Chain Monte Carlo algorithm (AM-MCMC). The AM-MCMC is adopted to produce a large number of weights for single models in an ensemble model. The quantitative forecasting (prediction) and its confidence interval at a given probability (90%) are obtained by multiplying the single-model predictions by the mean and the confidence interval of the weights assigned to those predictions, respectively. In this study, the annual precipitation (single annual value of each year) is adopted to test the performance of the new scheme. The precipitation of the forecast year is obtained from the precipitation forecast of the previous p years (p is the autocorrelation order of annual precipitation series). The results show that the new scheme for probabilistic forecasting for precipitation has better forecasting accuracy than single-model predictions; the RMSE is less than 139, and the MARE is less than 8.99%. Moreover, the new scheme for probabilistic forecasting for precipitation gets great probabilistic metrics, the CRPS ranges from 0.009 to 0.036, the reliability ranges from 0.001 to 0.008, and sharpness ranges from 24 to 77.

Automated summary

A new scheme for probabilistic precipitation forecasting is proposed in this study, which uses signal decomposition techniques and ensemble models to improve forecasting accuracy and probabilistic metrics compared to single-model predictions.