Water Quality Prediction Based on Machine Learning Using Multi-Dimension Input LSTM

Water quality plays a crucial role in maintaining ecosystem health, relying on physicochemical indicators like Dissolved Oxygen (DO). The effectiveness of Long Short-Term Memory (LSTM) networks in predicting water quality timeseries with large datasets and long temporal sequences has been well-established. However, previous research utilizing LSTM with Single-Dimension (1-D) input may limit predictive accuracy for complex data sets. This paper proposes the utilization of LSTM with Multi-Dimension (MD) Inputs to enhance performance and increase accuracy. Additionally, Moving Average (MA) is incorporated in the preprocessing stage to improve accuracy by smoothing the data. The study utilizes water quality data from the LOIS project, comprising approximately 24,000 data points with four features for each location. The results revealed that the incorporation of MD inputs led to higher R-squared values, with a 1.32% increase at location ID (FID) 29040 and a 0.03% increase at FID 1584376. Furthermore, the MD model demonstrated the significantly lower Mean Absolute Error (MAE) and Mean Squared Error (MSE) values, specifically 3.54E-07 at FID 29040 and 2.42E-05 at FID 1584376. These values were found to be lower by 95.34% and 33.52%, respectively, compared to the 1-D approach. Therefore, the predicted accuracy of LSTM using MD inputs surpassed that of the 1-D approach.