A Sensitive LSTM Model for High Accuracy Zero-Inflated Time-Series Prediction

The prevalence of zero values in zero-inflated time-series (ZI-TS) data poses significant challenges for traditional LSTM networks in learning long-term dependencies and trends. Specifically, the high proportion of zeros dilutes the influence of non-zero values, leading LSTM model to frequently predict zeros, which ultimately reduces prediction accuracy. A Weighted Zero-inflated Sensitive LSTM model (WZS-LSTM) which integrates both a stacked LSTM architecture and a Weighted Zero-inflated Sensitive (WZS) loss function was proposed to improve LSTM's prediction accuracy in ZI-TS data. First, the stacked LSTM structure enhances the model's capacity to capture long-term dependencies. Second, the model dynamically adjusts its focus towards non-zero values through the WZS loss function. Thorough experiments were conducted on UCI time-series dataset "WeatherAUS" and "Population", and four typical time-series prediction model Prophet, ARIMA, LSTM, and the hurdle model are selected for comparative analysis of algorithm performance, while "WeatherAUS" is a zero-inflated time-series data, and "Population" is a normal time-series data. The results demonstrated that the WZS-LSTM model improved prediction accuracy, reducing errors by at least 2.38% when compared with the selected models: Prophet, ARIMA, LSTM, and the hurdle model in ZI-TS data, while the predictive performance of WZS-LSTM for general time-series data was comparable to ARIMA, Prophet and traditional LSTM. In addition, WZS reduced the error by an additional 0.21% compared to the best-performing loss function. Finally, it was demonstrated on real-world datasets that WZS-LSTM offers significant advantages in predicting ZI-TS data and holds great potential for broader application.