Abstract: In view of the non-stationarity and complexity of coupling features of daily water supply time series, a random forest model based on scale feature fusion (SF-RF) was constructed by incorporating wavelet decomposition technique and random forest model. Firstly, the raw time series with a single scale was decomposed into multi-scale subsequences with both low and high frequencies using discrete wavelet transformation. Secondly, the multi-scale feature in each subsequence was simulated using the random forest model. Finally, the predicted value was obtained by linear fusion using the sub-results in each scale. Features in the highest frequency scale did not participate in the forecast. Compared with single RF model, feed-forward neural network (FFNN) and fusion model SF-FFNN, the proposed SF-RF model has the highest correlation coefficient 0.913 and the lowest normalized root mean square error 0.056, indicating that the proposed model has the highest forecasting accuracy and can be utilized for daily water supply forecasting.

Key words: daily water supply, wavelet transform, random forest, forecasting model, scale feature