Pond dissolved oxygen prediction method and system
Through the dual-branch structure and gated fusion module of the DBEG network model, the problem of model complexity and computing resource requirements in pond dissolved oxygen prediction is solved, efficient dissolved oxygen prediction is achieved, and prediction accuracy and efficiency are improved.
Patent Information
- Application Number
- CN202411081793.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing technology faces the problems of high model complexity and high computing resource requirements in pond dissolved oxygen prediction. Especially when processing large-scale data, the traditional attention mechanism has high computing cost and high memory consumption, making it difficult to effectively predict complex space-time dependencies. time series of dissolved oxygen.
Using the DBEG network model, the BiLSTM and EffAttn module are combined with the dimensional feature branch and time branch, and the data is integrated using the gated fusion module to reduce the computational complexity and improve prediction efficiency. The sliding window mechanism and mutual information feature selection are used to build an improved attention mechanism. two-branch model for prediction.
It improves the prediction accuracy and efficiency of dissolved oxygen in the pond, reduces the computational complexity and memory consumption, realizes more efficient data processing and prediction results, and demonstrates superior performance in complex data environments.
Smart Images

Figure CN119090050A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of pond dissolved oxygen prediction, and in particular to a pond dissolved oxygen prediction method and system. Background Art
[0002] The importance of dissolved oxygen in aquaculture cannot be ignored. It is one of the key factors to maintain the ecological balance of water bodies and ensure the healthy growth of aquatic organisms. The level of dissolved oxygen directly affects the respiration of fish and other aquatic organisms in the aquaculture water, and thus affects their growth rate, immunity and survival rate. Therefore, monitoring and predicting the dissolved oxygen content in the aquaculture environment is of great practical significance for improving aquaculture efficiency and reducing aquaculture losses.
[0003] The complexity of the aquaculture environment poses a great challenge to the prediction of dissolved oxygen content. As a special time series, dissolved oxygen data is not only affected by natural factors, such as seasonal changes, sunshine duration, temperature fluctuations, etc., but also by human intervention, such as the adjustment of breeding density and changes in feed input. The combined effect of these factors makes the dissolved oxygen time series data highly nonlinear, with large-scale fluctuations and complex spatiotemporal dependence, which brings unprecedented challenges to the time series prediction of dissolved oxygen.
[0004] In recent years, many scholars have devoted themselves to the research of dissolved oxygen prediction. Although the current research has achieved good results in dissolved oxygen prediction, it still faces the challenges of high model complexity and large computing resource requirements when processing large-scale data. At the same time, although the traditional attention mechanism has achieved good results in time series prediction, it still has the problems of high computational cost and large memory consumption. Summary of the invention
[0005] The purpose of the present invention is to provide a method and system for predicting dissolved oxygen in a pond, which can improve the prediction accuracy and efficiency of dissolved oxygen in a pond.
[0006] To achieve the above object, the present invention provides the following solutions:
[0007] A method for predicting dissolved oxygen in a pond, comprising:
[0008] Get the input sequence given by the prediction task of the target pond;
[0009] The input sequence is input into the trained DBEG network model, and the sequence prediction is performed by the DBEG network model to obtain the pond dissolved oxygen prediction result; the DBEG network model includes a dimensional feature branch and a time branch; wherein the dimensional feature branch includes a first EffAttn module and a first BiLSTM module connected in sequence; the time branch includes a second EffAttn module and a second BiLSTM module connected in sequence; the first BiLSTM module and the second BiLSTM module are also connected to a gated fusion module.
[0010] Optionally, the training process of the DBEG network model is:
[0011] Acquire training data; the training data includes a training sequence and a corresponding prediction result;
[0012] Construct a dual-branch model based on an improved attention mechanism;
[0013] The training data is input into the dual-branch model, the mean square error between the prediction sequence output by the model and the prediction result is calculated, and back propagation is performed through the Adam optimizer to update the network parameters to obtain a trained model, and the trained model is determined as the DBEG network model.
[0014] Optionally, the input sequence is input into a trained DBEG network model, and sequence prediction is performed by the DBEG network model to obtain a pond dissolved oxygen prediction result, specifically including:
[0015] Inputting the dimensional feature sequence and the time feature sequence in the input sequence into the dimensional feature branch and the time branch respectively for attention calculation to obtain dimensional features and time series features;
[0016] The BiLSTM module is used to further extract the long-term dependencies between the dimensional features and the temporal features, and the dynamic information between the data is integrated to obtain the final dimensional features and temporal features;
[0017] The final dimensional features and temporal features are input into the gated fusion module and fused according to the assigned weights to obtain the pond dissolved oxygen prediction results.
[0018] Optionally, the training data is collected by:
[0019] The collected pond water quality data set is grouped and segmented to adapt to the model's requirements for data format; first, the mutual information method is used as the basis for feature selection, and the time series length and prediction sequence length in each group of data are intercepted according to the selected dimensional features. These three data correspond to three parts in each group of data: dimensional feature sequence, time series and prediction sequence; the sliding window mechanism is used for grouping, and the window length is the sum of the time series length and the prediction sequence length. The window moves one bit each time, that is, there is only one bit difference between two adjacent groups of data; after completing the data grouping, 80% of the data groups are intercepted as training data sets and 20% of the data groups are used as verification data sets.
[0020] Optionally, the time series is obtained by transposing the dimensional feature sequence.
[0021] Optionally, the gated fusion module adopts a Gate module.
[0022] Optionally, the EffAttn module is composed of a linear layer and a Softmax function, and the linear layer performs a simplified linear transformation to directly weight the input data.
[0023] The present invention also provides a pond dissolved oxygen prediction system, comprising:
[0024] A task determination unit, used to obtain an input sequence given by a prediction task of a target pond;
[0025] The pond dissolved oxygen prediction unit is used to input the input sequence into the trained DBEG network model, perform sequence prediction through the DBEG network model, and obtain the pond dissolved oxygen prediction result; the DBEG network model includes a dimensional feature branch and a time branch; wherein the dimensional feature branch includes a first EffAttn module and a first BiLSTM module connected in sequence; the time branch includes a second EffAttn module and a second BiLSTM module connected in sequence; the first BiLSTM module and the second BiLSTM module are also connected to a gated fusion module.
[0026] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0027] The present invention discloses a method and system for predicting dissolved oxygen in a pond. The method comprises obtaining an input sequence given by a prediction task of a target pond; inputting the input sequence into a trained DBEG network model, performing sequence prediction through the DBEG network model, and obtaining a prediction result of dissolved oxygen in the pond; the DBEG network model comprises a dimensional feature branch and a time branch; wherein the dimensional feature branch comprises a first EffAttn module and a first BiLSTM module connected in sequence; the time branch comprises a second EffAttn module and a second BiLSTM module connected in sequence; the first BiLSTM module and the second BiLSTM module are also connected to a gated fusion module. The present invention can improve the prediction accuracy and efficiency of dissolved oxygen in a pond. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0029] Figure 1 It is a schematic flow chart of the method for predicting dissolved oxygen in a pond of the present invention;
[0030] Figure 2 is the overall structure diagram of the DBEG network model in this embodiment;
[0031] Figure 3 : is a structural diagram of the EffAttn module in this embodiment;
[0032] Figure 4 is a result diagram of the ablation experiment performed in this embodiment;
[0033] Figure 5 This is a curve diagram of the prediction results of each model when compared with different models in this embodiment. DETAILED DESCRIPTION
[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] The purpose of the present invention is to provide a method and system for predicting dissolved oxygen in a pond, which can improve the prediction accuracy and efficiency of dissolved oxygen in a pond.
[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0037] like Figure 1-Figure 2 As shown, the present invention provides a method for predicting dissolved oxygen in a pond, comprising:
[0038] Step 100: Obtain an input sequence given by the prediction task of the target pond;
[0039] Step 200: Input the input sequence into the trained DBEG network model, perform sequence prediction through the DBEG network model, and obtain the pond dissolved oxygen prediction result; the DBEG network model includes a dimensional feature branch and a time branch; wherein the dimensional feature branch includes a first EffAttn module and a first BiLSTM module connected in sequence; the time branch includes a second EffAttn module and a second BiLSTM module connected in sequence; the first BiLSTM module and the second BiLSTM module are also connected to a gated fusion module.
[0040] As a specific implementation method, the specific processing flow of the above steps is as follows:
[0041] Step 1: Data preprocessing to obtain training data set and validation data set.
[0042] The collected pond water quality data set is grouped and segmented to adapt to the model's requirements for data format; first, the mutual information (MI) method is used as the basis for feature selection, and the historical sequence length and predicted sequence length in each group of data are intercepted according to the selected dimensional features. These three data correspond to three parts in each group of data: dimensional feature sequence, historical sequence, and predicted sequence; a sliding window mechanism is used for grouping, and the window length is the sum of the historical sequence length and the predicted sequence length. The window moves one bit each time, that is, there is only one bit difference between two adjacent groups of data; after completing the data grouping, 80% of the groups of data are intercepted as training data sets and 20% of the groups of data are used as validation data sets.
[0043] Step 2: With the help of the training data set obtained in step 1, 32 groups of training data are randomly selected each time, the dimensional feature sequence is transposed to obtain the time feature sequence, and input into the EffAttn framework. The time feature sequence and dimensional feature sequence in each group of data are respectively input into the EffAttn framework model of the time branch and the dimensional feature branch. The two branches extract the time series features and dimensional features of the input sequence respectively through the EffAttn framework.
[0044] like Figure 2As shown in the figure, input represents input features, Positional Encoding represents position encoding, EffAttn represents attention calculation module, BiLSTM represents bidirectional long short-term memory network, Gate represents gating module, Flatten&Linear represents flattening operation, Linear represents fully connected layer, and output represents output.
[0045] like Figure 3 As shown, input represents the input feature ( Figure 2 The input in represents the features of the original features after mutual information feature selection. Figure 3 The input in Figure 2 In the input, the feature after position encoding represents the input of EffAttn), Linear represents the linear mapping operation, Softmax represents the normalization operation, and output represents the output of EffAttn.
[0046] Figure 3 The overall structure of the EffAttn of the present invention is shown. The EffAttn framework consists of a linear layer and a Softmax function. The input of the linear layer is the historical sequence of the original data after feature engineering in step 1. The time feature sequence and dimension feature sequence in each group of data are respectively input into the EffAttn modules of the time branch and the dimension feature branch. The two branches extract the time series features and dimension features of the input sequence respectively through the EffAttn module. EffAttn directly weights the input data by simplifying the linear transformation, thereby reducing the complex matrix operations involved in the traditional QKV framework.
[0047] Specifically, EffAttn avoids the need for comprehensive pairwise comparison of all sequence elements, and instead calculates the attention distribution in a more efficient way, greatly reducing the computational complexity. This simplified approach reduces the number of model parameters and computation time, reducing the computational complexity from the original square of the sequence length to the linear level, namely O(n·d), O(n·d) is expressed as: reducing the computational complexity from the original O(n^2) to O(n·d), where n represents the length of the input sequence and d represents the feature dimension after linear transformation. The following are the specific steps of EffAttn:
[0048] Use a linear layer on the input X in Perform a linear transformation and then use the Softmax function for normalization. The overall formula of EffAttn is as follows:
[0049] W = softmax(Linear(X in ))
[0050] Among them, the formula represents the normalized linear processing, and the Linear(·) function represents the linear transformation operation.
[0051] Calculate the weighted time series output, and perform element-by-element multiplication of the attention weight A with the input data:
[0052] X out =Ae X in
[0053] Where X in and X out Represents input data and output of EffAttn respectively, A is X in The corresponding attention distribution weight, e is the Hadamard product, that is, element multiplication.
[0054] Step 3: Use BiLSTM to further extract long-term dependencies in the time series features and dimension features output by EffAttn and integrate dynamic information between data.
[0055] Furthermore, in step 3, the time series features and dimensional features output by EffAttn are used to further extract the long-term dependencies in the features using BiLSTM, and the dynamic information between the data is integrated. The final time series features and dimensional features are input into the Gate module. The time series features and dimensional features finally obtained in step 3 are fused through the Gate module, and different weights are assigned to them and then fused into one result, thereby obtaining the final generated prediction sequence.
[0056] Step 4: Input the final time series features and dimensional features into the Gate module. The Gate module fuses the time series features and dimensional features obtained in step 3, assigns different weights to them, and then fuses them into one result, thereby obtaining the final generated prediction sequence.
[0057] Furthermore, the gate control mechanism Gate described in step 4 is responsible for giving X F and X TAssign corresponding weights and merge the results of the attention mechanism on two different dimensions to produce the final prediction output. The core idea of the gating mechanism is to learn how to balance and integrate information from the two branches of the model through a simple neural network. First, the outputs of the two branches are connected to form a new input Z with a dimension of 2T×u, where T represents the length of the sequence or the number of time steps, and u represents the feature dimension or size of the input at each time step. Then use a fully connected layer to learn how to assign weights to the two branches through a neural network. This can be achieved through a fully connected layer with a Softmax activation function to ensure that the sum of the output weights is 1. Finally, use these weights to weight the merge of the outputs of the two branches. The overall formula of the Gate mechanism is as follows:
[0058] Z=[X T X F ]
[0059] W = softmax(Linear(Z))
[0060] Y=X T ·W T +X F ·W F
[0061] W=[W T W F ]
[0062] Among them, X F represents the output of the feature branch, X T Indicates the output of the time branch, [] indicates that X F and X T Splicing is performed according to the time dimension, W is the learned weight distribution, W F and W T is the corresponding X extracted from the weight W F and X T The weight part of .
[0063] For the above weight formula, W = softmax (Linear (Z)) represents the weight obtained after normalizing the input Z linearly, where Z is X T and X F The concatenated matrix, and W = [W T W F ] indicates X T and X F The weight after the concatenation of the corresponding weights.
[0064] Step 5: Calculate the mean square error (MSE) between the final generated prediction sequence and the prediction sequence obtained in step 4, and then perform back propagation through the Adam optimizer to update the network parameters; finally, the trained DBEG network model is obtained.
[0065] Furthermore, the mean square error (MSE) described in step 5 and the root mean square error RMSE, mean absolute error MAE and determination coefficient R described in step 9 are 2 The formula is as follows:
[0066]
[0067] Where n is the number of samples, y i is the true value, is the predicted value, is the average of the true values.
[0068] Step 6: Test the DBEG network model using the validation dataset;
[0069] Input the validation data set obtained in step 1 into the trained DBEG network model, and finally obtain the prediction sequence generated based on the validation data set;
[0070] Step 7: Calculate the mean square error (MSE) between the prediction sequence generated based on the validation data set and the true prediction sequence, calculate the mean square error (MSE) of all groups of data, and then average them to obtain the MSE error based on the validation data set;
[0071] Step 8: Repeat steps 2 to 7 until the mean square error (MSE) obtained in step 7 no longer decreases, indicating that the model performance cannot be improved. The network parameters are updated and the model training ends.
[0072] Step 9: Input the input sequence given by the prediction task into the trained model obtained in step 8, perform sequence prediction, output the final prediction sequence, complete the prediction, and calculate the test set evaluation indicators root mean square error RMSE, mean absolute error MAE and determination coefficient R 2 .
[0073] In this embodiment, the dual branch is divided into a feature branch and a dimension branch. The main purpose of the feature branch is to calculate the attention for the feature dimension of the data. However, the calculation based on the feature dimension alone may ignore the mutual dependence between the time steps in the time series. In order to make up for this, a position encoding module is introduced to capture the relative position information of different time steps in the time series, thereby providing the model with the sequence context of the time step. The position encoding module encodes the positional relationship of different time steps in the output X of the dual branch structure part, and obtains the positional encoding PE (Positional Encoding) corresponding to X. PE reflects the positional relationship of each time step in the entire X. Then PE is added to X to obtain the new data Xpe, which not only contains the original feature information, but also integrates the time position information, and then sends it to the EffAttn module for attention calculation of the feature dimension. In order to further improve the model's processing ability for long sequences, BiLSTM is introduced after the EffAttn module. BiLSTM can effectively capture and learn long-term dependencies in time series data. By forward and backward analysis of the sequence, it can extract and integrate dynamic information between time steps from two directions. In this way, through the processing of the BiLSTM layer, the feature branch can not only use the current feature information, but also comprehensively consider the context information of the entire time series, greatly improving the accuracy and reliability of the prediction. F .
[0074] The overall design of the time branch is similar to that of the feature branch. It is also sent to the BiLSTM to calculate the output X of the branch after the attention calculation of the EffAttn module. T . Unlike the feature branch, the time branch performs attention calculation on the time dimension of X, and can understand and learn the position information of different time steps in the time series, so there is no need for the position encoding module to extract the position information in X for it. However, EffAttn will calculate the last dimension of the data by default, that is, for the input multivariate dissolved oxygen time series, it will perform attention calculation on the feature dimension. In order to realize EffAttn's calculation of the time dimension, a transposition-attention calculation-reduction calculation process is designed. The calculation process formula is as follows:
[0075] X trans =transpose(X)
[0076] X weighted = transpose(EffAttn(X trans )
[0077] Among them, X trans represents the features of the original input after transposition, Xweighted represents the calculation result of the attention weight, the transpose(·) function represents the transposition calculation operation, and EffAttn is the name of the module composition in the dimension feature branch and the time branch.
[0078] The transposed data is sent to the EffAttn module, which will trans The last dimension of the time dimension is used for attention calculation, and then transposed again. The time branch finally generates the output result X T With the output X of the feature branch F Send them into the Gate together.
[0079] Table 1 shows the experimental results of eight methods, DBEG, CNN-BiLSTM-Attention, CNN-Attention, WT-LSTM, GRU, MLP, XGBoost, and SVR, on the dissolved oxygen dataset under the same experimental conditions. The measurement criteria are root mean square error RMSE, mean absolute error MAE, and determination coefficient R 2 . Under specific experimental conditions, the experimental results of the best performing models are bolded in the table. As can be seen from the table, the overall prediction error of the last 6 deep learning models is better than that of the two machine learning models, SVR and XGBoost, which proves the advantage of deep learning time series prediction models in processing non-stationary dissolved oxygen data. At the same time, the prediction effect of the last 4 relatively complex deep learning models is better than that of the two relatively simple deep learning models, MLP and GRU, indicating that dissolved oxygen data requires more complex models for learning and prediction. Finally, it can be seen that DBEG has the best evaluation indicators in 3 aspects, and the prediction curve is also the closest to the true value, which proves the ability of DBEG in dissolved oxygen prediction.
[0080] Table 1 Experimental results on dissolved oxygen dataset
[0081]
[0082] Figure 4 The experimental results of the ablation experiment conducted under the same experimental conditions to evaluate the contribution of each component in the model to the overall performance of the model are shown. In order to prove the rationality and effectiveness of the main modules and methods involved in the DBEG proposed in this invention, three variants of DBEG are designed. Each model variant only deletes or changes a certain module and retains other modules to make it the same as the entire model. The specific details are as follows:
[0083] DBEG-StdAttn means replacing the EffAttn of both branches in DBEG with the standard attention mechanism. DBEG-TimeBranch means removing the feature branch and retaining only the time branch, a model variant with a single branch structure. DBEG-FeatureBranch means removing the time branch and retaining only the feature branch, a model variant with a single branch structure. DBEG-NoGate means replacing Gate with a model variant with direct addition.
[0084] Figure 5 The results of dissolved oxygen prediction of each model in the comparative experiment in Table 1 are shown. The prediction performance of the SVR model shows significant deficiencies compared with other models, especially the prediction deviations at multiple key points are more obvious. Relatively speaking, CNN-Attention also has certain prediction errors in a few cases. The MLP, WT-LSTM and CNNBiLSTM-Attention models perform well in overall prediction. Although there are slight errors in some specific areas, these errors are relatively small and do not affect their overall prediction accuracy. The DBEG model shows excellent ability in overall prediction performance. Its prediction curve has the highest degree of fit with the actual dissolved oxygen curve and the smallest prediction error, which proves its strong prediction ability.
[0085] Beneficial effects of the present invention:
[0086] The dual-branch structure adopted by the DBEG model embodies a methodology for comprehensive analysis of time series data. This structure not only improves the model's depth of understanding of the data, but also enhances its ability to capture complex data patterns. Through the parallel processing of feature branches and time branches, the DBEG model can simultaneously grasp the temporal dynamics and feature correlations in the time series, which is crucial for accurately predicting environmental indicators such as dissolved oxygen. When integrating the outputs of the two branches, the introduction of Gate is another major innovation in model design. It not only serves as a bridge connecting the feature and time branches, but also optimizes the information integration process by dynamically adjusting the weights of the outputs of each branch. This mechanism ensures the flexibility and adaptability of the model when facing different data characteristics, further improving the accuracy and reliability of predictions.
[0087] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0088] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A method for predicting dissolved oxygen in a pond, characterized in that: include: Get the input sequence given by the prediction task of the target pond; The input sequence is input into the trained DBEG network model, and the sequence prediction is performed by the DBEG network model to obtain the pond dissolved oxygen prediction result; the DBEG network model includes a dimensional feature branch and a time branch; wherein the dimensional feature branch includes a first EffAttn module and a first BiLSTM module connected in sequence; the time branch includes a second EffAttn module and a second BiLSTM module connected in sequence; the first BiLSTM module and the second BiLSTM module are also connected to a gated fusion module.
2. The method for predicting dissolved oxygen in a pond according to claim 1, wherein: The training process of the DBEG network model is: Acquire training data; the training data includes a training sequence and a corresponding prediction result; Construct a dual-branch model based on an improved attention mechanism; The training data is input into the dual-branch model, the mean square error between the prediction sequence output by the model and the prediction result is calculated, and back propagation is performed through the Adam optimizer to update the network parameters to obtain a trained model, and the trained model is determined as the DBEG network model.
3. The pond dissolved oxygen prediction method according to claim 1, characterized in that: The input sequence is input into the trained DBEG network model, and the DBEG network model performs sequence prediction to obtain the pond dissolved oxygen prediction result, which specifically includes: Inputting the dimensional feature sequence and the time feature sequence in the input sequence into the dimensional feature branch and the time branch respectively for attention calculation to obtain dimensional features and time series features; The BiLSTM module is used to further extract the long-term dependencies between the dimensional features and the temporal features, and the dynamic information between the data is integrated to obtain the final dimensional features and temporal features; The final dimensional features and temporal features are input into the gated fusion module and fused according to the assigned weights to obtain the pond dissolved oxygen prediction results.
4. The method for predicting dissolved oxygen in a pond according to claim 2, characterized in that: The method for collecting the training data is as follows: The collected pond water quality data set is grouped and segmented to adapt to the model's requirements for data format; first, the mutual information method is used as the basis for feature selection, and the time series length and prediction sequence length in each group of data are intercepted according to the selected dimensional features. These three data correspond to three parts in each group of data: dimensional feature sequence, time series and prediction sequence; the sliding window mechanism is used for grouping, and the window length is the sum of the time series length and the prediction sequence length. The window moves one bit each time, that is, there is only one bit difference between two adjacent groups of data; after completing the data grouping, 80% of the data groups are intercepted as training data sets and 20% of the data groups are used as verification data sets.
5. The method for predicting dissolved oxygen in a pond according to claim 4, characterized in that: The time series is obtained by transposing the dimensional feature sequence.
6. The method for predicting dissolved oxygen in a pond according to claim 1, characterized in that: The gated fusion module adopts a Gate module.
7. The method for predicting dissolved oxygen in a pond according to claim 1, characterized in that: The EffAttn module is composed of a linear layer and a Softmax function, and the linear layer performs a simplified linear transformation to directly weight the input data.
8. A pond dissolved oxygen prediction system, characterized in that: include: A task determination unit, used to obtain an input sequence given by a prediction task of a target pond; The pond dissolved oxygen prediction unit is used to input the input sequence into the trained DBEG network model, perform sequence prediction through the DBEG network model, and obtain the pond dissolved oxygen prediction result; the DBEG network model includes a dimensional feature branch and a time branch; wherein the dimensional feature branch includes a first EffAttn module and a first BiLSTM module connected in sequence; the time branch includes a second EffAttn module and a second BiLSTM module connected in sequence; the first BiLSTM module and the second BiLSTM module are also connected to a gated fusion module.
Citation Information
Patent Citations
Method and apparatus for predicting dissolved oxygen in aquaculture
CN107016453A
Radar echo extrapolation forecasting method and system and storage medium
CN117665825A
Water body dissolved oxygen content prediction method, related device and storage medium
CN118072856A
Cited By
Petrochemical harbor district safety risk analysis method based on deep learning
CN119296302A
