Page 79 - AJWEP-22-5

Hybrid optimization for LSTM DO prediction

A B

C D

Figure 2. Scatter plots of water quality parameters versus dissolved oxygen. (A) pH versus dissolved oxygen;
(B) Temperature versus dissolved oxygen; (C) Turbidity versus dissolved oxygen; (D) Conductivity versus
dissolved oxygen.

features into a multi-feature predictive model allows study developed a prediction model based on an LSTM
for the exploration of both individual effects and their network and employed the permutation feature importance
potential. Advanced machine learning models–such as method to evaluate the importance of input features.
27
Random Forest, XGBoost, and LSTM–are particularly This method involves independently shuffling the time
well-suited for this task, as they can automatically series of each feature and observing the change in model
capture complex linear and nonlinear dependencies performance (measured by MSE), thereby assessing the
among features. Therefore, by leveraging multi- actual impact and contribution of each feature to the
26
feature integration and the learning capabilities of prediction outcomes within the specific model.
these algorithms, we anticipate improved accuracy in During the evaluation, the baseline MSE of the model
forecasting DO levels. on the original test set was 0.027023. Subsequently, five
repeated random shuffling experiments were conducted
4.1.2. Characteristic importance analysis for each feature. The average increase in MSE and its
To gain a deeper understanding of the contribution of standard deviation were calculated as indicators of
various water quality indicators in predicting DO, this feature importance. The results are shown in Figure 3.

Volume 22 Issue 5 (2025) 73 doi: 10.36922/AJWEP025210165