Hybrid Distribution Separation for Prediction of Heterogeneous Zero-Inflated Time Series
ZhiXin Huang, Jiaxiang Lin, Qian-Qian Chen, Z. Zhang · 2023
Weather forecasting is an important application of time series predictions that is widely used in agriculture, fisheries, and the maintenance of power and communications network infrastructure. A variety of time series analyzing approaches were proposed by the researchers, which perform well in the prediction by mining the temporal characteristics of data. However, most of them ignore the zero-inflated features and heterogeneous distributions of the time series, thus the accuracy of prediction is usually unsatisfactory. To solve this problem, a hybrid distribution separation model (called HDS) is proposed to discriminate and handle different distributions that existed in the zero-inflated data for a better prediction accuracy. First, a preprocessing classification model is adopted to partition the zero and non-zero observations, thus achieving the effect of eliminating the zero-inflated phenomenon. Then, a three-layer Stacked LSTM is applied to deal with the mixture feature of different distributions, which can increase the complexity of the model and make the approximation function more effective, hence a better accuracy is obtained during the prediction than single-layer LSTM. Finally, several comparative experiments are performed on the icing data with different model evaluation indicators, which is collected in Fuzhou City during 2018 and 2020. The experimental results demonstrate that, for the icing data with slightly changed, the HDS model has an RMSE of 0.2249, which is 37.6% lower than the best compared algorithm Prophet (whose RMSE value is 0.3604). For the icing data with sharply changed, the HDS model has an RMSE value of 0.2421, which is slightly higher than the best compared algorithm Prophet (whose RMSE value is 0.1076), but it is lower than the other methods. Overall, the proposed HDS model is superior to the compared time series prediction methods in RMSE.