Short-Term Prediction of Groundwater Level Based on Spatiotemporal Correlation

Abstract

Groundwater is an important component of water resources. Excessive exploitation can lead to water scarcity, while excessive groundwater can cause problems such as terrain collapse. Therefore, it is crucial to obtain timely information on groundwater levels (GWL) and make predictions accordingly. Previous methods, considering only the time series of the sites did not consider the spatial relationships between the sites due to the difficulty to handle unstructured site location data. This paper proposes a GWL prediction model based on GCN-LSTM, each GWL site is regarded as a node of the graph. The spatial topological relationship between the sites is determined through Pearson correlation coefficient and Euclidean distance, and the adjacency matrix is constructed. Using the graph convolutional neural network (GCN) mines the spatial characteristics between groundwater level sites, and then inputs the time series data with spatial characteristics into the long short-term memory neural network (LSTM) to mine the temporal characteristics to predict groundwater. In this paper, the GWL data from 18 sites were selected to predict GWL in the next three day. The mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE) are used as evaluation criteria, the mean absolute error of the predicted GWL was 0.13 m, the mean square error was 0.02 m, and the root mean square error was 0.14 m, the proposed prediction model has lower errors than the LSTM and TGCN models. Compared with the long short-term memory network model, this model can shorten the training time and improve the prediction efficiency. And it can reflect the dynamic changes of GWL at multiple sites more efficiently and obtain the changing trend of GWL in the entire region, which has certain practical application value.