Coupling the Causal Inference and Informer Networks for Short-term Forecasting in Irrigation Water Usage

Precise and reliable irrigation water use (IWU) prediction is beneficial for irrigation district reservoirs interaction and water resources management. However, existing methods face the challenges of high prediction errors at extreme points and accumulative error problem. Meanwhile, ignoring the effects of spurious relationships are among the driving factors on prediction results. This study introduces the Peter and Clark Momentary Conditional Independence (PCMCI) causal inference method to analyze driving factors. The causal inference results of PCMCI are taken as input to the IWU prediction model. This investigation constructs six IWU forecasting models, including the Informer neural network, long short-term memory (LSTM) neural network, attention-based LSTM network, the Prophet model, random forest model, and seasonal autoregressive integrated moving average (SARIMA). The performance of these six models and their variants are evaluated and compared by the long-term month IWU data series of the Zhanghe irrigation area of China. The results show that the Informer model based on self-attention mechanism is more advantageous than others. The PCMCI method can overcome the spurious relationships, and contribute to a clearer understanding of physical mechanisms, as compared to the correlation analysis. Combined with the dynamic time warping barycenter averaging (DBA) data augmentation method, the proposed DBA-PCMCI-Informer method can reduce the prediction errors at extreme points, and improve the IWU forecasting accuracy. This approach alleviates the accumulative error problem, enabling high accuracy even in multistep prediction.