A hybrid deep learning framework for real-time makes predictions in aquaculture

Performance Standards

During the experiment, several models were used to predict the results of the dataset. Several standard indices were used to assess the accuracy of predictions for these models and to facilitate comparisons between them. These include the mean absolute error (MAE), MSE, RMSE, and R². These indices allow for a quantitative assessment of the accuracy and effectiveness of predictions generated by each operating model. They are defined as follows:

Simulation results and analysis

All models were trained with the same time series data to predict future dissolved oxygen levels. Figures 6 and 7 show the results of DO level estimation using the model. This indicates that all models can estimate data trends with reasonable accuracy. Analysis of both figs. 6 and 7 reveal that the combined model is superior to the individual model, and that the proposed CNN-SA-Bisru produces the best results (called zoom-in set). Single models, such as the GRU and SVR models, showed lease agreements between high volatility and actual curves.

Figure 6 shows that the fit curves for the CNN-Bisru combination model exhibit a higher degree of match and fit than the fit curves of a single Bisru model, and are generally closer to the actual curve. CNN feature extraction reduced overlaps and predictive properties from time series data.

Figure 8 shows the violin plots for all predictive models. The violin plot shows the data distribution and allows for comparison of multiple groups. The white dots in the center of the violin plot indicate the median value, while the width indicates the density of the data. From the shape of each violin and the median position in Figure 8, we can see that the median values of all predictive models move up and down. The model is more symmetrical, indicating that the predicted values are concentrated. The model proposed in this study has the most consistent shape with the original data, indicating that the model has the best predictive performance and can obtain good predictive results with the data distribution.

The CNN-SA-Bisru model is closest to the original data and is the best. Figure 9 shows the ablation performance of the proposed model. For a single Bisru model, only parallel operations can be achieved that are suitable for capturing long data dependencies. However, Bisru alone lacks emphasis on important information, so the results of model predictions are the most deviant from the original data. The CNN-Bisru model adds CNN to feature extraction. This allows you to add the advantages of CNN to Bisru to better interpret your data. CNN feature extraction reduced overlaps and predictive properties from time series data. Therefore, the predictive performance of this model increases. The auto-articulation method adjusted the weights assigned to the data obtained by the CNN function. The newly calibrated weight distribution was then entered into Bisru to determine which parameter combinations provided the optimal results for CNN-SA-Bisru. The experimental findings revealed that the model has excellent generalization capabilities in terms of its ability to predict water quality metrics for widely cultivated water bodies. Therefore, this model can produce optimal effects even in peaks and troughs.

Table 2 shows the quantitative results for all predictive models. Dissolved oxygen concentrations in water quality parameters for intensive aquaculture are predicted to produce the best results combined with the CNN-SA-Bisru prediction model suggested in this study. MSE, MAE, RMSE, and r² It was 0.0022, 0.0341, 0.0471, and 0.9765, respectively. For the individual models, the root mean root error for Bisru is 0.0822, which is 0.58%, 0.54%, 2.4%, and 4.92%, smaller than CNN, LSTM, GRU, and SVR, respectively.

This study compared the CNN-Bisru and Bisru models. The rating metrics averaged MSE, MAE, RMSE, and R². The CNN-Bisru model has 1.64%, 0.22%, 1.41% lower MAE, MSE, and RMSE, respectively, with a high R R² 3.75% compared to the Bisru model. Enhancements can be credited to Bisru models that receive raw time series data directly without undergoing CNN-based feature extraction. As a result, this increased data redundancy and computational effort, and ultimately affected the accuracy of the model. By incorporating CNN feature extraction before entering the data into the BISRU model, the original input data can be extracted for hierarchical information. This improves the predictive performance of the model, resulting in better results.

The CNN-SA-Bisru model has 1.06%, 0.24%, 2.1% lower MAE, MSE, and RMSE, respectively, with a high R R² 2.05% compared to the CNN-Bisru model. This improvement may be attributed to the efficiency of one-dimensional convolution of CNNs when extracting features from original time series data. However, this approach is limited because all hidden layer units have the same weight. This increases computational complexity and reduces the effectiveness of the model. The Auto-Joint (SA) method assigns various importance to elements in the hidden layer. This provides a greater emphasis on essential data, minimizes record wear and slightly improves prediction accuracy. After a different weighted SA mechanism, the data is entered into the Bisru model. Compared to models without SA mechanisms, performance has been significantly improved. The predicted value has slight errors in the measurement, resulting in the optimal predictive effect.

The findings from the above experiments reveal that integration of CNN and auto-joint function improves the accuracy of the Bisru model. Due to improved parameters, this proposed approach has excellent predictive accuracy and adaptability.

The prediction framework introduced by LAP has produced promising results¹. Nevertheless, our study utilized a clear dataset, chose a unique feature selection approach, and adopted a different prediction technique tailored to the problem. Compared to the LAP model, the approach in this study shows excellent prediction accuracy. Their models use built-in feature selection. This automatically assesses the importance of functionality without manually setting thresholds or using domain knowledge to reduce component dimensions⁴⁰. However, this feature selection method has its drawbacks. It is limited to representational features that properly capture complex image or data features.

In contrast, in our study, extracting CNN features can learn more expressive features through convolution and pooling operations, and thus achieve some degree of accuracy improvement.⁴¹. Regarding the selection of predictive models, this reference model selects an ML (machine learning)-based model, similar to the models in this paper. This reference prediction model determines the RF (Random Forest). This has excellent descriptive and interpretable features, higher parallelism, and faster training⁴²However, because the dependencies between the data could not be captured, the prediction accuracy cannot be improved, and the prediction accuracy cannot be improved. In contrast, the predictive model Bisru in this paper further improves prediction accuracy and robustness, as it increases data dependence and allows better functional learning to be performed.

Furthermore, Figure 10 shows the error profiles for the eight predictive models. Figure 10(a). The GRU shows high volatility in the error range, indicating that the model's ability to generalize to the sample is unstable and not sufficiently adaptive. Figure 10(b). The range of variation for the LSTM error values is more stable than the range of errors for the GRU. However, it increases in some regions, indicating that model performance deteriorates when dealing with certain sequential data. Still, there is room for improvement. Figure 10(c). CNN shows that the error rate fluctuates sideways with increasing sample, indicating that the model has limited adaptation to new data. Figure 10(d). The overall trend for SVR errors is relatively flat. However, certain points have spikes. This is due to the insensitive model to sample changes, leading to inaccurate predictions of individual samples. Figure 10(e). The volatility of Bisru errors is lower than the volatility of errors in GRU and LSTM models, indicating that the model achieves a constant balance of stability and adaptability and has better predictive performance. Figure 10(f). Although CNN-Bisru combines the benefits of CNN and Bisru, the volatility of error rates suggests that there may be room for adjustments to model fusion. Figure 10(g). CNN-SA-BILSTM shows that BILSTM is complicated by the structure of the prediction model, which reduces the predictive performance of the model and does not encourage prediction of water quality parameters dissolved oxygen. Figure 10(h). CNN-SA-Bisru shows that inclusion of autocatalytic mechanisms is beneficial to improve overall model performance. These results show that the CNN-SA-Bisru model is most effective in predicting dissolved oxygen content. Using this method, researchers can obtain more accurate predictions of dissolved oxygen content for environmental monitoring and water quality control.

The combined CNN, autoarticulation, and Bisru model appears as the main method for estimating dissolved oxygen levels in water metrics, essential for high density aquaculture. This study shows that the recommended model can calibrate accuracy and effectiveness and successfully capture variations in dissolved oxygen concentration. As a result, it responds to real-world demands and support to make important decisions about the quality of water used in intensive agriculture. Compared to CNN, Bisru, LSTM, GRU, and SVR models, the CNN-SA-Bisru model is better than predicting do intensive agriculture. Based on these results, it is reasonable to assume that the proposed model may be used for various environmental monitoring and prediction tasks.