A dynamic stratified prediction method for water temperature in high-level pond aquaculture

Abstract

The present invention provides a method for dynamic hierarchical prediction of water temperature based on high-level pond aquaculture, wherein high-level pond water parameters including at least water temperature data and air temperature data and external meteorological data are input into a multimodal fusion adaptive collaborative network, and a dynamic merging method is used to segment the time signal set into time segments; and a prediction model TimeFusionNet fused with a progressive attention mechanism is used to perform time series prediction to obtain prediction results. The present invention provides a set of targeted water temperature prediction schemes for high-level pond aquaculture environments, which can be widely used in high-level pond aquaculture environments, provide a refined water temperature control scheme for the aquaculture industry, and improve aquaculture efficiency. At the same time, the scheme can also provide important data support for environmental monitoring and ecological research, and has high application value and social benefits.

Description

Technical Field

[0001] The invention belongs to the technical field of prediction methods or systems, and in particular relates to a water temperature dynamic stratification prediction method based on high-level pond aquaculture. Background Art

[0002] As an efficient aquaculture method, high-level pond aquaculture has been widely used around the world. In the high-level pond aquaculture environment, water temperature is a key indicator that reflects and affects the physiological functions and growth efficiency of aquaculture organisms, and its control and prediction are particularly important. Experiments show that the change in water temperature in high-level ponds is mainly due to environmental factors such as sunshine, air temperature, wind speed, and air pressure. However, at the same time, the response of water temperature changes at different depths to environmental factors is different; at the same depth, the response to environmental factors at different times is also different. In high-level ponds, water temperature changes have strong temporal and dynamic characteristics and are affected by multiple factors. Traditional static clustering methods cannot automatically adjust clustering results according to changes in time series, resulting in the inability to adapt to the law of water temperature changes in a timely manner. At the same time, the water temperature of different water layers is significantly different, and the temperature changes between water layers have a high degree of temporal and spatial dependence. Existing prediction methods usually fail to effectively consider the differences in water temperature in water layers, and it is difficult to provide accurate predictions for each water layer. At present, the monitoring and prediction of water temperature in high-level pond aquaculture mainly rely on manual observation and single sensor measurement, resulting in the inability to respond to water temperature changes in a timely manner.

[0003] In addition, although traditional prediction methods can provide certain prediction capabilities, they are mostly focused on a single model or variable, often ignoring the micro-environmental changes inside the aquaculture pond, such as the impact of air temperature, wind speed, air pressure and water temperature in different layers on the water layer temperature. It is difficult to meet the needs of high-precision prediction of water layer temperature combined with the unique environmental factors of high-level ponds. They are not adaptable to the specific environment of high-level ponds and have limitations in practical applications. In summary, the existing methods lack highly targeted hierarchical prediction technology, fail to achieve multi-dimensional and accurate modeling of different water layers, water bodies and environmental conditions, and cannot achieve personalized water temperature prediction based on water layer characteristics. Summary of the invention

[0004] In view of this, in view of the defects and shortcomings of the existing technology and actual needs, the present invention aims to propose a dynamic stratified prediction scheme of water temperature based on high-level pond aquaculture. By integrating the Internet of Things system, data processing algorithm and machine learning model, high-precision prediction of water temperature in high-level pond aquaculture can be achieved, in order to provide scientific decision-making support for aquaculture managers.

[0005] It aims to provide a scientific temperature control solution for aquaculture through stratified analysis and time series prediction of water temperature in the spatiotemporal dimension. The present invention combines high-level pool water quality parameters and external meteorological data to provide a multimodal fusion adaptive collaborative network, which uses a dynamic merging method to segment the time signal set into time segments. And the prediction model TimeFusionNet fused with the independently designed progressive attention mechanism is used to realize time series prediction, effectively predict the water temperature changes at each depth, and efficiently process and analyze the spatiotemporal dynamic characteristics of water temperature changes. Furthermore, the optimized gray wolf algorithm improved by the multi-guided return strategy is used to optimize the hyperparameters of the prediction model TimeFusionNet. As a preferred embodiment, a specially designed multi-node, multi-depth Internet of Things data collector design is also provided.

[0006] The technical solution specifically adopted by the present invention to solve the technical problem is:

[0007] A dynamic hierarchical prediction method for water temperature based on high-level pond aquaculture: high-level pond water parameters including at least water temperature data and air temperature data and external meteorological data are input into a multimodal fusion adaptive collaborative network, and a dynamic merging method is used to segment the time signal set into time segments; and a prediction model TimeFusionNet fused with a progressive attention mechanism is used to perform time series prediction to obtain prediction results.

[0008] It should be noted that among the multimodal data used in the prediction scheme of the present invention, water temperature data and air temperature data are the most core necessary data. However, in order to achieve more accurate predictions, other meteorological data such as wind force, wind speed, and air pressure are generally also required.

[0009] Furthermore, the water temperature data is obtained by field collection of water data at different locations and depths in the high-level pool. The data obtained is collected continuously to form a continuous original data set. The original data set is preprocessed using a wavelet transform threshold denoising method, and different types of data are timestamped.

[0010] Furthermore, the multimodal fusion adaptive collaborative network is an adaptive collaborative network with a two-dimensional network structure, in which each node ( i , j ) are associated with a weight vector v ij , the dimension of the weight vector matches the feature dimension of the input data, that is, , where the subscripts 1, 2, ..., d refer to the dimensions.

[0011] Furthermore, the dynamic merging method is used to divide the time signal set into time segments based on a determined sliding window and sliding step size, and the input water temperature time signal is dynamically merged using an adaptive collaborative network, thereby obtaining the merged results of water temperatures at different depths; the time signals with the same merged results are merged into a time segment as the input of the next time series prediction model; the specific steps include:

[0012] After denoising, d The water temperature data is standardized to obtain a standardized time signal, and the time signal is divided into multiple time segments using a sliding window;

[0013] Using the random initialization method, for each node ( i , j ) v ij (0), whose elements Randomly select values ​​within the range of the water temperature data; randomly shuffle the input time segments to ensure the randomness and independence of the data during each training, and then select each training sample in turn Carry out subsequent processing;

[0014] For each training sample k , Manhattan distance is used to calculate the degree of association between the training sample and each node in the network; through calculation, the node with the smallest distance is found, which is defined as the core adaptation node and recorded as ( m , n );

[0015] Update the weights of network nodes; for nodes ( i , j ), whose weight vector v ij In the t+1 The round update formula is as follows:

[0016]

[0017] in, t represents the current training round, represents the learning rate, represents the influence radius, represents the adjustment function based on the position relationship and the range of influence; represents the initial learning rate, represents the attenuation rate, represents the initial influence radius, Indicates the decay rate;

[0018] Repeat the process of finding core adaptation nodes and updating node weights until the predetermined training rounds are reached or the convergence condition is met; the convergence condition formula is as follows:

[0019]

[0020] in, It indicates setting the average change threshold of node weights for several consecutive rounds. M and N It is the dimension of the network;

[0021] The time segments with the same merging results are merged into one time segment, and the data with similar water temperature characteristics are integrated to provide data input for the subsequent time series prediction model.

[0022] Furthermore, based on the merging results and merging windows determined by the multimodal fusion adaptive collaborative network, the water temperature of the same layer, the water temperature of some adjacent layers and meteorological data are input into the prediction model TimeFusionNet model for prediction, and multi-scale learning and differential features are combined to capture the dynamic changes in the time series, and the performance of the model in the time series prediction task is optimized:

[0023] The prediction model TimeFusionNet model introduces a progressive attention mechanism to learn features at different levels in stages, including the global learning stage, the local feature learning stage, and the micro-feature learning stage. The three stages extract three time steps in units of weeks, days, and minutes, respectively, and use the weighted summation method to weight the differential features and the original features:

[0024] λ t z =softmax ( W z , △ x t ), z= {1,2,3}

[0025] in, W z It is z The weight matrix of the learning stage, △ x t is the differential feature, λ t z is the attention weight of the z-th learning stage; z= {1,2,3} correspond to the global learning stage, local feature learning stage, and micro-feature learning stage, respectively;

[0026] Fusion of progressive attention mechanism and original features: For different time steps, after calculating the progressive attention weights at different stages, the differential features and the original features are fused to obtain the final input features As input prediction model TimeFusionNet model is trained:

[0027]

[0028] in, is the differential feature of the zth stage, The original feature.

[0029] Furthermore, the improved gray wolf optimization algorithm based on the multi-guided return strategy is used to optimize the parameters of the prediction model TimeFusionNet. The specific process includes:

[0030] Initialize the gray wolf population parameters, including the gray wolf population size and the maximum number of iterations t max , spatial dimension dim, upper and lower limits ub and lb of the search space, define RMSE as the fitness function of the optimization algorithm, that is, the optimization goal;

[0031] Calculate the individual fitness values ​​of the wolf pack and determine the wolf α , β , δ;

[0032] Update the parameters A and C of the added disturbance factor;

[0033]

[0034] Where: randn represents a random number that follows a standard normal distribution, represents the peak position of the disturbance factor, a is the convergence factor, r1 and r2 are two random numbers in [0,1];

[0035] Integrate the newly added disturbance factors into the updating process of wolf and prey positions;

[0036]

[0037] D α , D β , D δ express α , β , δ the distance between the wolf and other individuals, X α , X β , X δ Respectively α , β , δ is the current position of the wolf, X(t) is the current position of the individual gray wolf;

[0038]

[0039] Where: X(t+1) represents the position of the remaining gray wolves after being updated at the next moment under the influence of X1, X2, and X3; X1, X2, and X3 represent the positions of the remaining gray wolves under the influence of X1, X2, and X3 respectively. α , β , δ wolves, the remaining gray wolves α , β , δ is the step length and direction of the wolf's advance; A1, A2, A3 represent a set of dynamic coefficients;

[0040] Judgment condition t / t max Is it greater than rand? If so, use the multi-guided return strategy to calculate C. If the conditions are met, calculate the new position and perform return adjustment; otherwise, jump directly to the next step;

[0041] Where: S represents the reentry strength coefficient, is the prey position, d is the dimension;

[0042] Continue iterating until the maximum number of iterations t is reached max ;

[0043] Output the optimal solution α The position and fitness value of the wolf; when the fitness value is the smallest, that is, the RMSE reaches the minimum value, it means that the improved gray wolf optimization algorithm based on the multi-guided return strategy has completed the parameter optimization of the TimeFusionNet prediction model.

[0044] The training adopts the optimized results of the multimodal fusion adaptive collaborative network as model parameters; in feature extraction, the change of input features is calculated through dilated convolution and differential processing; the preprocessing of training normalizes the feature set including the original features and the differential features, and divides it into a training set and a test set.

[0045] Furthermore, the prediction model TimeFusionNet uses improved LSTM units as training units for each deep data; during the training process, the gradient is updated through a standard back-propagation algorithm; the improved LSTM unit uses RMSE as the fitness function of the optimization algorithm, and uses an improved grey wolf optimization algorithm based on a multi-guided return strategy to optimize the hyperparameters in LSTM; for input data of three time scales, a residual connection layer is established between the feature fusion layers of each time scale; for the LSTM units in the hidden layer, a combination of the Huber loss function and the contrast loss is used as the connection layer, and parallel LSTM units are used to ensure that features of different time scales can effectively collaborate.

[0046] Furthermore, as a preferred solution of the present invention, multiple Internet of Things data collectors are used to collect water data at different positions and depths of the high-level pool on the spot, including central nodes and edge nodes, and the number and layout positions are configured according to the size of the high-level pool; the central node is used to measure the central water temperature and environmental data of the high-temperature pool, and the edge node is used to measure the water temperature and environmental data of the peripheral heat exchange area; the data acquired by the Internet of Things data collector is uploaded to the Internet of Things platform via wireless communication.

[0047] Furthermore, the Internet of Things data collector includes a head controller, a pole multi-depth water temperature collector and a bottom counterweight block, the head controller includes a power supply module with a solar panel, a wireless communication module and a collection control circuit, and the external structure includes a waterproof shell, a lightweight float and a four-way anti-roll and drift wing; the pole multi-depth water temperature collector includes a plurality of temperature sensors arranged at intervals, and a sensor cascade node, and the external structure includes a cylindrical waterproof shell and a waterproof docking joint for cascading; the bottom counterweight block is connected to the pole multi-depth water temperature collector by a locking structure, which is used to keep the Internet of Things data collector vertically downward in the high-level pool.

[0048] And, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of a method for dynamic stratification prediction of water temperature based on high-level pond aquaculture as described above are implemented.

[0049] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for dynamically stratifying water temperature based on high-level pond aquaculture as described above.

[0050] In view of the defects and shortcomings of the prior art, the present invention and its preferred solution provide a set of targeted water temperature prediction solutions for high-level pond aquaculture environments, which can be widely used in high-level pond aquaculture environments, provide a refined water temperature control solution for the aquaculture industry, and improve aquaculture benefits. At the same time, the solution can also provide important data support for environmental monitoring and ecological research, and has high application value and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0052] Figure 1 It is an overall flow chart of the embodiment of the present invention;

[0053] Figure 2 This is a flowchart of the adaptive collaborative network workflow of an embodiment of the present invention;

[0054] Figure 3This is a flow chart of the improved gray wolf optimization algorithm based on the multi-guided return strategy according to an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of the arrangement of multi-node collectors in a high-level pool according to an embodiment of the present invention;

[0056] Figure 5 The figure is a schematic diagram of the structure of the Internet of Things data collector according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the features and advantages of the present invention more clearly understood, the following embodiments are specifically described in detail as follows:

[0058] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0059] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0060] like Figure 1 As shown, in the embodiment of the present invention, an innovative multimodal fusion adaptive collaborative network is provided in combination with high-level pool water quality parameters and external meteorological data, and a dynamic merging method is used to segment the time signal set into time segments. And the independently designed progressive attention mechanism fusion prediction model TimeFusionNet is adopted to realize time series prediction, effectively predict the water temperature changes at each depth, and efficiently process and analyze the spatiotemporal dynamic characteristics of water temperature changes. Furthermore, the optimized gray wolf algorithm improved by the multi-guided return strategy is used to optimize the hyperparameters of the prediction model TimeFusionNet. As a preferred embodiment, a specially designed multi-node, multi-depth Internet of Things data collector design is also provided.

[0061] In this embodiment, water data at different locations and depths of the high-level pool are first collected on-site.

[0062] As a preferred embodiment of this embodiment, the collected water temperature data and crawled meteorological data are collected continuously and form a continuous raw data set. The raw data set includes water temperature data (℃), air temperature data (℃), wind force data (level), wind speed data (km / h), air pressure data (hPa), atmospheric humidity data (%), air quality data ( / ), visibility data (km), etc.

[0063] The collected data can be preprocessed through the Internet of Things platform; this embodiment preferably uses a wavelet transform threshold denoising method to preprocess the water body data such as water temperature, water depth, time, etc. collected in the field and the meteorological data obtained by the crawler (for example, obtained from the China Meteorological Administration), and perform timestamp alignment.

[0064] Based on the determined sliding window and sliding step size, the input water temperature time signal is dynamically merged using an adaptive collaborative network to obtain the merged results of water temperatures at different depths. The time signals with the same merged results are merged into a time segment as the input of the next time series prediction model.

[0065] Then, as one of the core designs of this embodiment, an ACN adaptive cooperative network (Adaptive Cooperative Network) is used to merge the water temperature data at different depths in each sliding window, including:

[0066] After denoising, d The water temperature data are standardized to obtain a standardized time signal, and a sliding window is used to divide the time signal into multiple time segments.

[0067] Construct an adaptive collaborative network with a two-dimensional network structure, where each node ( i , j ) are associated with a weight vector v ij , the dimension of the weight vector matches the feature dimension of the input data, that is, .

[0068] Using the random initialization method, for each node ( i , j ) v ij (0), whose elements Randomly select values ​​within a range related to the reasonable range of water temperature data.

[0069] The input time segments are randomly shuffled to ensure the randomness and independence of the data in each training, and to avoid the model's dependence on the data order. Then, each training sample is selected in turn. Perform subsequent processing.

[0070] For each training sample k, the Manhattan distance is used to calculate its association with each node in the network. Through calculation, the node with the smallest distance (or the largest similarity) is found and defined as the core adaptation node, denoted as (m,n);

[0071] Update the weights of network nodes; for nodes ( i , j ), whose weight vector v ij In the t+1 The round update formula is as follows:

[0072]

[0073] in, t represents the current training round, represents the learning rate, represents the influence radius, Represents the adjustment function based on position relationship and influence range.

[0074] Repeat the above process of finding core adaptation nodes and updating node weights until the predetermined training round is reached or the convergence condition is met. The convergence condition formula is as follows:

[0075]

[0076] in, It indicates setting the average change threshold of node weights for several consecutive rounds. M and N are the dimensions of the network.

[0077] The time segments with the same merging results are merged into one time segment, and the data with similar water temperature characteristics are integrated to provide more representative and coherent data input for the subsequent time series prediction model.

[0078] The process of the above adaptive collaborative network is as follows Figure 2 shown.

[0079] Based on the merged results and merged windows determined in the above steps, the water temperature of the same layer, the water temperature of some adjacent layers, and the meteorological data are input into the TimeFusionNet model of the progressive attention mechanism for prediction. Specifically, this embodiment proposes a progressive attention mechanism PAM (Progressive Attention Mechanism), which aims to combine multi-scale learning and differential features to capture dynamic changes in time series and optimize the performance of the model in time series prediction tasks:

[0080] In terms of parameter selection, the results of ACN adaptive collaborative network optimization are used as model parameters. In terms of feature extraction, the change of input features is calculated through dilated convolution and differential processing.

[0081] The merged feature set (including original features and differential features) is normalized.

[0082] The dataset is divided into training set and test set in a ratio of 8:2.

[0083] The introduction of the progressive attention mechanism enables the model to learn features at different levels in stages. The process consists of three stages: global learning stage (coarse learning), local feature learning stage (medium granularity learning) and micro feature learning stage (fine learning). Correspondingly, the three stages extract three time steps in units of weeks, days and minutes, respectively, and use the weighted summation method to weight the differential features and original features. The specific formula is as follows:

[0084] λ t z =softmax ( W z , △ x t ), z= {1,2,3}

[0085] in, W z It is z The weight matrix of the learning stage, △ x t is the differential feature, λ t z is the attention weight of the z-th learning stage; z= {1,2,3} correspond to the global learning stage, local feature learning stage and micro feature learning stage respectively.

[0086] Fusion of progressive attention mechanism and original features: For different time steps, after calculating the progressive attention weights of different stages, the differential features and the original features are fused to obtain the final input features:

[0087]

[0088] in, is the differential feature of the zth stage, is the original feature, and finally the fused feature is obtained .

[0089] The fused features The TimeFusionNet prediction model is trained as input. As a preferred method, the model uses the improved LSTM unit as the training unit for each deep data. During the training process, the gradient is updated by the standard back-propagation algorithm. Compared with the traditional LSTM algorithm, there are three differences: First, the introduction of the progressive attention mechanism and differential features enables TimeFusionNet to model time series at multiple scales. Through different attention learning stages (global, local, and micro), the long-term trend, local change, and short-term fluctuation in the time series are captured respectively. Second, the hyperparameters in the LSTM unit are optimized. Specifically, a multi-guided reentry strategy is proposed to optimize the gray wolf algorithm, and RMSE is selected as the fitness function of the optimization algorithm. The gray wolf algorithm optimized based on the multi-guided reentry strategy is used to optimize the hyperparameters in the LSTM. Third, for the input data of three time scales, a residual connection layer is established between the feature fusion layer of each time scale; for the LSTM unit in the hidden layer, a combination of Huber loss function and contrast loss is used as the connection layer, and parallel LSTM units are designed to ensure that the features of different time scales can be effectively coordinated.

[0090] The above improved gray wolf optimization algorithm based on multi-guided return strategy is different from the traditional gray wolf algorithm in the following four points:

[0091] First, a dynamic perturbation factor and a reentry mechanism are introduced during the wolf pack search process to enhance the flexibility and exploration ability of the search process. By dynamically adjusting the convergence factor α during the search process of individual wolves, the factor gradually decreases with the increase in the number of iterations, thereby increasing the exploration range in the early stages of the search and promoting more accurate convergence in the later stages. Unlike the fixed linear decrease in the traditional gray wolf algorithm, the dynamically adjusted convergence factor enables the algorithm to better adapt to the optimization needs of different stages and improves the overall search efficiency.

[0092] Second, during the wolf pack search process, the disturbance factors A and C are introduced to increase the diversity of the search path, reduce the possibility of falling into the local optimum, and enhance the adaptive ability of the optimization algorithm.

[0093] 3. In the process of updating individual wolves, a new reentry strategy is introduced. The position update of each individual depends not only on the current search path, but also on the reentry strength coefficient S. This coefficient adjusts the individual position update according to the current search progress and the distance to the target location, allowing the wolf pack to jump more flexibly in the search space, thus avoiding the dilemma of falling into the local optimal solution.

[0094] Fourth, a dynamic fitness evaluation mechanism based on the current progress is additionally introduced. In each iteration, the algorithm will determine whether the current progress meets specific conditions. If so, the individual search path will be adjusted through the return mechanism.

[0095] Therefore, the improved gray wolf optimization algorithm based on the multi-guided return strategy is used to optimize the parameters of the TimeFusionNet prediction model, including:

[0096] Initialize the gray wolf population parameters, including the gray wolf population size and the maximum number of iterations t max , spatial dimension dim, upper and lower limits ub and lb of the search space, and define RMSE as the fitness function of the optimization algorithm, that is, the optimization objective.

[0097] Calculate the individual fitness values ​​of the wolf pack and determine the wolf α , β , δ;

[0098] Update the parameters A and C of the added disturbance factor;

[0099]

[0100] Where: randn represents a random number that follows a standard normal distribution, represents the peak position of the disturbance factor, a is the convergence factor, r1 and r2 are two random numbers in [0,1];

[0101] The newly added disturbance factors are integrated into the updating process of wolf and prey positions.

[0102]

[0103] D α , D β , D δ express α , β , δ the distance between the wolf and other individuals, X α , X β , X δ Respectively α , β , δ is the current position of the wolf, X(t) is the current position of the individual gray wolf;

[0104]

[0105] Where: X(t+1) represents the position of the remaining gray wolves after being updated at the next moment under the influence of X1, X2, and X3; X1, X2, and X3 represent the positions of the remaining gray wolves under the influence of X1, X2, and X3 respectively. α , β , δ wolves, the remaining gray wolves α ,β , δ is the step length and direction of the wolf's advance; A1, A2, A3 represent a set of dynamic coefficients;

[0106] Judgment condition t / t max Is it greater than rand? If so, use the multi-guided return strategy to calculate C. If the conditions are met, calculate the new position and perform return adjustment; otherwise, jump directly to the next step;

[0107] Where: S represents the reentry strength coefficient, is the prey position, d is the dimension;

[0108] Continue iterating until the maximum number of iterations t is reached max .

[0109] Output the optimal solution α The position and fitness value of the wolf. When the fitness value is the smallest, that is, the RMSE reaches the minimum value, it means that the parameter optimization of the TimeFusionNet prediction model has been completed using the improved gray wolf optimization algorithm based on the multi-guided return strategy.

[0110] The flowchart of the improved gray wolf optimization algorithm based on the multi-directional return strategy is as follows: Figure 3 shown.

[0111] Based on the above design, R 2 (coefficient of determination), MAE (mean absolute error), and RMSE (root mean square error) are used to evaluate the performance of the model proposed in this embodiment on the validation set.

[0112] In addition, considering that the acquisition of data of the scheme of the present invention requires on-site collection of water data at different positions and depths in the high-level pool, this embodiment also designs an optimal multi-node, multi-depth Internet of Things data collector, which is suitable for on-site collection of water data at different positions and depths in the high-level pool.

[0113] In the preferred solution of this embodiment, the multi-node, multi-depth IoT data collector is composed of a number of central nodes and edge node collectors, and the number and layout location can be configured according to the size of the high-level pool. The central node is used to measure the water temperature and environmental data in the center of the high-temperature pool, and the edge node is used to measure the water temperature and environmental data near the pool wall and other peripheral heat exchange areas. The data connection of each collector adopts wireless transmission and is configured on the upper computer software side. Each collector can work independently. The multi-node collector is arranged in the high-level pool as follows: Figure 4 shown.

[0114] A single IoT data collector consists of three parts: head: controller, rod: multi-depth water temperature collector, bottom: counterweight block, such as Figure 5As shown. The internal functions of its "head: controller" include solar panels, batteries and power circuits for providing energy consumption, data collectors for data acquisition, system control and analog-to-digital conversion processing, and wireless communication modules for communication. The external structure of its "head: controller" includes a waterproof shell, a lightweight float and a four-way anti-roll side drift wing. The internal functions of its "pole: multi-depth water temperature collector" include multiple temperature sensors with intervals of 10~20CM, and sensor cascade nodes that can be cascaded by 1 meter. The external structure of its "pole: multi-depth water temperature collector" includes a cylindrical waterproof shell and a waterproof joint cascaded by 1 meter. The "head: controller" can be cascaded in units of 1 meter length as needed, which is suitable for high-level pools of different depths. Its "bottom: counterweight block" consists of a lock structure that is stably connected to the bottom of the "pole: multi-depth water temperature collector" and a container for placing the counterweight block, which is used to make the device vertically downward. The data collected by the collector is uploaded to the Internet of Things platform via a wireless link.

[0115] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0116] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by the processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0117] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0118] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

[0119] The present invention is not limited to the above-mentioned optimal implementation mode. Anyone can derive various other forms of a dynamic stratification prediction method for water temperature based on high-level pond aquaculture under the inspiration of the present invention. All equal changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A method for dynamic stratification prediction of water temperature based on high-level pond aquaculture, characterized by: The high-level pool water parameters including at least water temperature data and air temperature data and external meteorological data are input into the multimodal fusion adaptive collaborative network, and the time signal set is segmented into time segments using the dynamic merging method; and the prediction model TimeFusionNet fused with the progressive attention mechanism is used to perform time series prediction to obtain the prediction results; The dynamic merging method is used to divide the time signal set into time segments based on a determined sliding window and sliding step size, and an adaptive collaborative network is used to dynamically merge the input water temperature time signal, thereby obtaining the merged results of water temperatures at different depths; the time signals with the same merged results are merged into a time segment as the input of the next time series prediction model; Based on the merging results and merging windows determined by the multimodal fusion adaptive collaborative network, the water temperature of the same layer, the water temperature of some adjacent layers and meteorological data are input into the prediction model TimeFusionNet model for prediction. Multi-scale learning and differential features are combined to capture the dynamic changes in the time series, and the performance of the model in the time series prediction task is optimized: The prediction model TimeFusionNet model introduces a progressive attention mechanism to learn features at different levels in stages, including a global learning stage, a local feature learning stage, and a micro-feature learning stage; The three stages extract three time steps in units of weeks, days, and minutes respectively, and all use the weighted summation method to weight the differential features and original features: λ t,z =softmax ( W z , △ x t )， z= {1,2,3} in, W z It is z The weight matrix of the learning stage, △ x t is the differential feature, λ t,z is the attention weight of the z-th learning stage; z= {1,2,3} correspond to the global learning stage, local feature learning stage, and micro-feature learning stage, respectively; Fusion of progressive attention mechanism and original features: For different time steps, after calculating the progressive attention weights at different stages, the differential features and the original features are fused to obtain the final input features As input prediction model TimeFusionNet model is trained: in, is the differential feature of the zth stage, is the original feature; The improved Grey Wolf Optimization Algorithm based on the multi-guided return strategy is used to optimize the parameters of the prediction model TimeFusionNet. The prediction model TimeFusionNet uses improved LSTM units as training units for each deep data; during the training process, the gradient is updated through a standard back-propagation algorithm; the improved LSTM unit uses RMSE as the fitness function of the optimization algorithm, and uses an improved grey wolf optimization algorithm based on a multi-guided return strategy to optimize the hyperparameters in the LSTM; for input data of three time scales, a residual connection layer is established between the feature fusion layers of each time scale; for the LSTM units in the hidden layer, a combination of the Huber loss function and the contrast loss is used as the connection layer, and parallel LSTM units are used to coordinate features of different time scales.

2. A method for dynamic stratification prediction of water temperature based on high-level pond aquaculture according to claim 1, characterized in that: The water temperature data is obtained by field collection of water data at different locations and depths in the high-level pool. The data obtained is collected continuously to form a continuous original data set. The original data set is preprocessed using a wavelet transform threshold denoising method, and different types of data are time-stamped.

3. A method for dynamic stratification prediction of water temperature based on high-level pond aquaculture according to claim 1, characterized in that: The multimodal fusion adaptive collaborative network is an adaptive collaborative network with a two-dimensional network structure, in which each node (i, j) is associated with a weight vector v ij , the dimension of the weight vector matches the feature dimension of the input data, that is, , where the subscripts 1, 2, ..., d refer to the dimensions.

4. A method for dynamic stratification prediction of water temperature based on high-level pond aquaculture according to claim 3, characterized in that: The dynamic merging method is used to divide the time signal set into time segments based on a determined sliding window and sliding step size, and an adaptive collaborative network is used to dynamically merge the input water temperature time signal, thereby obtaining the merged results of water temperatures at different depths; The specific steps of merging the time signals with the same merging results into one time segment as the input of the next time series prediction model include: After denoising, d The water temperature data is standardized to obtain a standardized time signal, and the time signal is divided into multiple time segments using a sliding window; Using the random initialization method, the weight vector of each node (i,j) is v ij (0), whose elements Randomly select values ​​within the range of the water temperature data; randomly shuffle the input time segments to ensure the randomness and independence of the data during each training, and then select each training sample in turn Carry out subsequent processing; For each training sample k , Manhattan distance is used to calculate the degree of association between the training sample and each node in the network; through calculation, the node with the smallest distance is found, which is defined as the core adaptation node and recorded as ( m , n ); Update the weights of the network nodes; for node (i,j), its weight vector v ij In the t+1 The round update formula is as follows: in, t represents the current training round, represents the learning rate, represents the influence radius, represents the adjustment function based on the position relationship and the range of influence; represents the initial learning rate, represents the attenuation rate, represents the initial influence radius, Indicates the decay rate; Repeat the process of finding core adaptation nodes and updating node weights until the predetermined training rounds are reached or the convergence condition is met; the convergence condition formula is as follows: in, It indicates setting the average change threshold of node weights for several consecutive rounds. M and N It is the dimension of the network; The time segments with the same merging results are merged into one time segment, and the data with similar water temperature characteristics are integrated to provide data input for the subsequent time series prediction model.

5. The method for dynamic stratification prediction of water temperature based on high-level pond aquaculture according to claim 1 is characterized in that: The specific process of optimizing the parameters of the prediction model TimeFusionNet by using the improved gray wolf optimization algorithm based on the multi-guided return strategy includes: Initialize the gray wolf population parameters, including the gray wolf population size and the maximum number of iterations t max , spatial dimension dim, upper and lower limits ub and lb of the search space, define RMSE as the fitness function of the optimization algorithm, that is, the optimization goal; Calculate the individual fitness values ​​of the wolf pack and determine the wolf α , β , δ; Update the parameters A and C of the added disturbance factor; Where: randn represents a random number that follows a standard normal distribution, represents the peak position of the disturbance factor, a is the convergence factor, r1 and r2 are two random numbers in [0,1]; Integrate the newly added disturbance factors into the updating process of wolf and prey positions; D α , D β , D δ express α , β , δ the distance between the wolf and other individuals, X α , X β , X δ Respectively α , β , δ is the current position of the wolf, X(t) is the current position of the individual gray wolf; Where: X(t+1) represents the position of the remaining gray wolves after the next round of training update, which is affected by X1, X2, and X3; X1, X2, and X3 represent the positions of the remaining gray wolves after the next round of training update, respectively. α , β , δ wolves, the remaining gray wolves α , β , δ is the step length and direction of the wolf's advance; A1, A2, A3 represent a set of dynamic coefficients; Judgment condition t / t max Is it greater than randn? If so, use the multi-guided return strategy to calculate C. If the conditions are met, calculate the new position and perform return adjustment; otherwise, jump directly to the next step; Where: S represents the reentry strength coefficient, is the prey position, d is the dimension; Continue iterating until the maximum number of iterations t is reached max ; Output the optimal solution α The position and fitness value of the wolf; when the fitness value is the smallest, that is, the RMSE reaches the minimum value, it means that the improved gray wolf optimization algorithm based on the multi-guided return strategy has completed the parameter optimization of the TimeFusionNet prediction model.

6. A method for dynamic stratification prediction of water temperature based on high-level pond aquaculture according to claim 1, characterized in that: Multiple IoT data collectors are used to collect water data at different locations and depths of the high-level pool, including central nodes and edge nodes, and the number and layout positions are configured according to the size of the high-level pool; the central node is used to measure the central water temperature and environmental data of the high-temperature pool, and the edge node is used to measure the water temperature and environmental data of the peripheral heat exchange area; the data obtained by the IoT data collector is uploaded to the IoT platform via wireless communication.

7. A method for dynamic stratification prediction of water temperature based on high-level pond aquaculture according to claim 6, characterized in that: The Internet of Things data collector includes a head controller, a multi-depth water temperature collector on the pole and a bottom counterweight. The head controller includes a power supply module with a solar panel, a wireless communication module and a collection control circuit, and the external structure includes a waterproof shell, a lightweight float and a four-way anti-roll and drift wing; the multi-depth water temperature collector on the pole includes a plurality of temperature sensors arranged at intervals, and a sensor cascade node, and the external structure includes a cylindrical waterproof shell and a waterproof docking joint for cascading; the bottom counterweight is connected to the multi-depth water temperature collector on the pole through a locking structure, which is used to keep the Internet of Things data collector vertically downward in the high-level pool.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of a method for dynamic stratification prediction of water temperature based on high-level pond aquaculture as described in any one of claims 1 to 5 are implemented.

Images