*3.3. Model Result Output*

After wavelet transform of DO, CODMn, TP, and NH3-N data, the decomposed lowfrequency wavelet coefficient *cA* and high-frequency wavelet coefficient *cD* were used as

the input of LSTM. Meanwhile, in order to verify that the ANN-WT-LSTM model had a higher prediction accuracy than other models, we selected ANN-LSTM, ARIMA, and NAR neural network models for comparison. The parameter settings of the other models are shown in Table 5.


**Table 4.** Wavelet reconstruction error table for each parameter.

**Table 5.** Parameter settings of ANN-WT-LSTM and traditional LSTM models.


<sup>a</sup> NAR neural network parameter settings: autoregressive coefficient lag = 3, number of hidden layer units = 300. <sup>b</sup> Set to reduce the learning rate by multiplying by a factor of 0.2 after 125 rounds of training. <sup>c</sup> BPNN parameter settings: number of hidden layer units = 300 and number of iterations = 250; <sup>d</sup> SSA-LSTM parameters: dim is 4, the range of learning rate is (0.001,1), the range of the maximum number of iterations is (10,500), number of sparrows = 5, warning value ST = 0.6, proportion of discoverers PD = 0.6, proportion of sparrows aware of danger SD = 0.2, and population size = 5; <sup>e</sup> ISSA-BPNN: based on the fact that the sparrow search algorithm tends to fall into a local optimum, a tent mapping was used to initialize the population for the sparrow search algorithm. <sup>f</sup> The parameters were set as follows: lower boundary (lb) of the weight threshold is −5 and upper boundary (ub) of the weight queue is 5.

The results of the ANN-WT-LSTM model are shown in Figures 9–12.

**Figure 9.** Prediction results and error images of the ANN-WT-LSTM model for high-frequency and low-frequency parts of the Do verification set.

**Figure 10.** Prediction results and error images of the ANN-WT-LSTM model for high-frequency and low-frequency parts of the CODMN verification set.

**Figure 11.** Prediction results and error images of the ANN-WT-LSTM model for high-frequency and low-frequency parts of the TP verification set.

**Figure 12.** Prediction results and error images of the ANN-WT-LSTM model for high-frequency and low-frequency parts of the NH3-N verification set.

The prediction error images of the ANN-LSTM model on the test set data and the comparisons with the original images are shown in Figures 13–16.
