Method for controlling maximum power of wave energy based on super-dimensional calculation and deep learning

Abstract

The invention discloses a wave energy maximum power control method based on super-dimensional calculation and deep learning, and the method comprises the steps: obtaining wave information in real time through a sensor, preprocessing the wave information, and inputting the preprocessed wave information into a hybrid prediction model for future wave situation prediction; according to the model, a three-layer super-dimensional calculation module (for enhancing the high-dimensional data processing capability of waves), a three-layer LSTM network (for capturing the dynamic characteristics of a wave time sequence), a CNN feature extraction layer (for analyzing the spatial distribution characteristics of the waves) and a novel dual-channel attention mechanism (for accurately focusing the key characteristics of the waves) are fused. According to the improved firework weighted average optimization algorithm, a dynamic mapping relation between explosion spark and Gaussian spark is constructed, a three-stage development strategy and a Levy flight exploration strategy are combined, and intelligent parameter optimization is specially carried out for maximum power tracking control of a wave energy device. The algorithm adopts a self-adaptive search space adjustment mechanism, so that the wave energy capture efficiency is remarkably improved, and premature convergence is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of marine renewable energy power generation equipment, and particularly relates to a maximum power control method for wave energy based on hyperdimensional computing and deep learning. Background Art

[0002] As an important carrier for distributed offshore energy supply, ocean wave energy demonstrates unique value in emerging fields such as deep-sea and far-sea resource development and ecological monitoring network construction. Currently, energy capture devices generally adopt a multi-field coupling technology architecture, and typical designs include three mainstream technical routes: the phase coordination mechanism of a float group, the fluid kinetic energy conversion system of an impeller matrix, and the hydraulic pendulum pressure drive device. Engineering practice shows that affected by the dynamic wave excitation and the frequency-domain adaptation deviation of the energy capture system, the steady-state energy conversion rate of existing devices mostly remains in the range of 30%-38%, which has become a key bottleneck restricting the development of the industry.

[0003] Traditional frequency-domain optimization theory constructs a control model based on the assumption of linear regular waves and uses a parameter adaptive adjustment mechanism to achieve quasi-static matching between the energy capture system and wave excitation. However, in the actual ocean environment, wave parameters exhibit significant non-Gaussian characteristics, and their wave height-period joint distribution has spatio-temporal multi-dimensional correlations, resulting in a double dilemma for traditional control methods: First, the wave distortion effect causes the instantaneous wave steepness to exceed the linear theory threshold, generating high-order harmonic energy components and leading to model structure mismatch; second, the prediction algorithm based on differential equation iteration has a dimension explosion, far exceeding the computing capacity of real-time control systems and making it difficult to meet the time constraints of real-time control.

[0004] The calculation delay of the traditional dynamic programming algorithm in a 4D state space reaches 2-5 s, exceeding the 0.1-0.5 s time window required by the control period. The root mean square error between the prediction result based on the simplified model and the real wave excitation exceeds 25%, leading to the risk of resonance frequency unlocking. In view of the contradiction between the dynamic matching accuracy and the calculation timeliness mentioned above, it is urgent to develop a new type of intelligent prediction control paradigm to achieve optimal energy capture under complex sea conditions. Summary of the Invention

[0005] Aiming at the technical problems of insufficient prediction accuracy, poor dynamic adaptability of control strategies, and weak environmental robustness in the above traditional wave energy control technologies, and the low power capture efficiency of wave energy devices in complex marine environments due to the strong coupling of wave spatio-temporal characteristics and significant non-stationary fluctuations, this technical solution provides a maximum power control method for wave energy based on hyperdimensional computing and deep learning. By constructing a hybrid architecture that synergizes hyperdimensional computing and deep learning, deeply integrating high-dimensional data processing, time-series dynamic modeling, spatial feature extraction, and attention focusing mechanisms, it breaks through the feature representation limitations of a single model. At the same time, a fireworks weighted average algorithm with dynamic mapping and adaptive boundary adjustment is designed to solve the problems of local convergence and response hysteresis in the traditional control parameter optimization process. Finally, the maximum power tracking control of wave energy devices under changing marine conditions is realized, improving the energy conversion efficiency and system operation stability, and effectively solving the above problems.

[0006] The present invention is realized through the following technical solutions:

[0007] A maximum power control method for wave energy based on hyperdimensional computing and deep learning, comprising the steps of:

[0008] S1: Obtain the wave information of the external waves through the sensors of the wave energy power generation device, including wave height and period;

[0009] S2: Preprocess the obtained wave information and output the preprocessed wave information;

[0010] S3: Input the wave information into the improved neural network prediction model and perform prediction to obtain the future wave information; the improved neural network prediction model includes three layers of hyperdimensional computing models, three layers of LSTM models, one layer of CNN model, and one layer of attention mechanism model; this multi-modal fusion architecture realizes the hierarchical deep mining and collaborative optimization of wave spatio-temporal characteristics; the prediction steps of the improved neural network prediction model include:

[0011] S3.1: Parallelly distribute the input data to three hyperdimensional computing modules for high-dimensional feature encoding;

[0012] S3.2: Input the hyperdimensional feature vectors output by each module into the LSTM models of the corresponding channels for wave time-series dynamic modeling;

[0013] S3.3: After the time-series features output by the three LSTM channels are fused through the feature splicing layer, they are uniformly input into the CNN module for analyzing the wave spatial distribution characteristics;

[0014] S3.4: Input the multi-scale spatial features extracted by the CNN into the spatial-channel dual attention mechanism, and strengthen the representation ability of key wave features through adaptive weight allocation, and finally output a high-precision wave situation prediction result;

[0015] S4: Use the improved fireworks weighted average algorithm to perform locking control on the wave energy power generation device.

[0016] Furthermore, the attention mechanism model includes a spatial attention model and a channel attention model; the output feature map of the CNN model is equally divided into two parts, which are respectively fed into the spatial attention module and the channel attention module for parallel processing; the spatial attention module and the channel attention module operate in parallel. The spatial attention model focuses on capturing the key information of the image or signal in the spatial dimension, and the channel attention model focuses on identifying and analyzing the important features in the channel or time dimension.

[0017] Furthermore, the processing flow of the spatial attention module is as follows: First, perform global average pooling on the input to capture spatial statistical information, then perform feature transformation through a convolutional layer, then perform batch normalization operation, and finally generate a spatial feature weight map through the Sigmoid function.

[0018] Furthermore, after the channel attention module compresses the spatial dimension by using global average pooling, it realizes linear transformation of the channel dimension through a fully connected layer, applies the ReLU activation function to enhance the non-linear expression ability, and finally generates a channel feature weight map through the Sigmoid function to complete the modeling of the interdependence between channels.

[0019] Furthermore, the specific operation method of step S4 is as follows:

[0020] S4.1: Initialize N different positions using the improved fireworks weighted average algorithm, and obtain explosion fireworks and Gaussian fireworks by setting off fireworks at N different positions. Each set-off firework represents the optimal locking time in the locking control situation at the previous moment. The explosion fireworks and Gaussian fireworks represent the single optimal locking time and the global optimal locking time at the previous moment respectively.

[0021] S4.2: Set the maximum number of iterations MAX it , and start calculating the weighted average. Rearrange all the locking times according to the two optimal locking times at the previous moment to obtain the number N c of the selected candidate solutions, and calculate the position of the weighted average. The formula is as follows:

[0022]

[0023] In the formula, nP is the total number, it is the current loop count, and then obtain the sum of the fitness values of all candidate solutions:

[0024]

[0025] Where, Sum Fitness is the sum of all fitness values, X i is the i-th candidate solution, Fitnes is the function for calculating fitness, and then the weighted average position X Miu is obtained as follows:

[0026]

[0027] S4.3: During the iteration process, the exploration and exploitation phases are defined by the following formula:

[0028]

[0029] In this work, K1 is the transformation function, α is a random constant, rand is a random value between 0 and 1, the threshold of K1 is taken as 0.5. When K1≥0.5, it moves according to the exploitation ability, while when K1<0.5, it moves according to the exploration ability;

[0030] S4.4: There are three strategies in total for the exploitation phase. The first one focuses on exploiting the search space, and the search space X i (it + 1) mainly depends on the current weighted average position and the personal best position X PersonBest obtained by the fireworks algorithm and the global optimal position X GlobalBest and is determined as follows:

[0031]

[0032] Where, w11, w12, and w13 are random values between 0 and 1, used to adjust the expansion of the search space;

[0033] The second strategy finally focuses on exploiting the search space established between the weighted average position of the entire population and the personal best position Under this strategy, the global optimal position is ignored. Compared with Strategy 1, this strategy reduces the search space and the search time, and the formula is as follows:

[0034]

[0035] Where, w21 and w22 are random values between 0 and 1;

[0036] Strategy three focuses on the search space established between the weighted average position and the global optimal position, and the formula is as follows:

[0037]

[0038] Where, w31 and w32 also have a value range between 0 and 1. Compared with Strategy 1, Strategy 2 and Strategy 3 mainly improve the convergence speed and accuracy;

[0039] S4.5: When K1 < 0.5, the system is in the exploration stage and is mainly optimized through two exploration strategies. The formula for Strategy 1 is as follows:

[0040]

[0041] S is the step size of Lévy flight, which is a random walk pattern with heavy-tailed characteristics. Its step size follows the Lévy distribution and is used in the optimization algorithm to enhance the global search ability; X GlobalBest,j (it) is the j-th position of the global optimal solution at the first iteration, and X i,j (it + 1) is the j-th position of the i-th solution at the (it + 1)-th iteration;

[0042] In some cases, the global optimal value obtained by the algorithm in the region around the global optimal position is too different from the ideal value of the global optimal position. In this case, the weighted average algorithm is at risk of converging to a local optimum. To solve this problem, a second strategy is adopted to adjust the search space, and the formula is as follows:

[0043] X i (it + 1) = rand * (UB min -LB min ) + LB min (9)

[0044] In the formula, LB min and UB min are the minimum values of the upper and lower bounds of the search space. Through this strategy, the search agent will move to a new position to find a better region;

[0045] S4.6: Finally, check the boundary conditions and update the optimal position of the single-body latching and the global optimal position at the current time;

[0046] S4.7: After calculating the optimal latching position and time, maximize the wave through latching control.

[0047] Beneficial effects

[0048] A wave energy maximum power control method based on hyperdimensional computing and deep learning proposed by the present invention has the following beneficial effects compared with the prior art:

[0049] (1) The present invention can achieve the intelligent prediction advantage of multi-dimensional feature fusion: through the innovative fusion architecture of hyperdimensional computing and deep learning, it breaks through the limitations of traditional single-modal analysis. Hyperdimensional computing enhances the structured representation ability of high-dimensional wave data, the LSTM network accurately captures the dynamic evolution law of wave periods, the CNN extracts multi-scale spatial fluctuation features, and the spatial-channel dual attention mechanism is used to focus on key information. This cross-dimensional collaboration significantly improves the spatio-temporal correlation of wave situation prediction, providing forward-looking input with high credibility for control decisions.

[0050] (2) The present invention can achieve a dynamic adaptive intelligent optimization mechanism: the improved fireworks weighted average algorithm constructs a dynamic mapping system between explosion fireworks and Gaussian fireworks, and realizes the collaborative optimization of local fine search and global rapid exploration through a three-stage development strategy. The exploration strategy guided by Lévy flight expands the search dimension, and the adaptive boundary adjustment mechanism effectively avoids local optimal traps, enabling the locking control parameters of the wave energy device to be dynamically optimized, and significantly improving the energy capture efficiency and system response agility.

[0051] (3) The present invention can achieve the improvement of system robustness under complex working conditions: the hybrid architecture design endows the algorithm with strong adaptability to ocean environment fluctuations. The hyperdimensional computing layer enhances the feature stability under noise interference, and the dual attention mechanism alleviates the influence of data drift through dynamic weight allocation. The optimized algorithm can correct the control strategy in real time to ensure stable power output under non-stationary working conditions such as sudden changes in wave height and disordered periods, providing a reliable guarantee for the engineering application of ocean energy devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is the overall process schematic diagram of the present invention.

[0053] Figure 2 is the architecture diagram of the neural network prediction model in the present invention.

[0054] Figure 3 is the architecture diagram of the attention mechanism in the present invention.

[0055] Figure 4 is the architecture diagram of the improved weighted average optimization algorithm in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Without departing from the design concept of the present invention, various modifications and improvements made by those of ordinary skill in the art to the technical solutions of the present invention should fall within the protection scope of the present invention.

[0057] Example 1:

[0058] A maximum power control method for wave energy based on hyperdimensional computing and deep learning, comprising the steps of:

[0059] S1: Obtain the wave information of the external waves through the sensors of the wave energy power generation device, including wave height and period;

[0060] S2: Preprocess the obtained wave information and output the wave information obtained after preprocessing;

[0061] S3: Input the wave information into the improved neural network prediction model and perform prediction to obtain future wave information;

[0062] The improved neural network prediction model innovatively integrates hyperdimensional computing technology, long short-term memory network (LSTM), convolutional neural network (CNN) and attention mechanism.

[0063] The improved neural network prediction model includes three layers of hyperdimensional computing models, three layers of LSTM models, one layer of CNN model and one layer of attention mechanism model; this multimodal fusion architecture realizes hierarchical deep mining and collaborative optimization of wave spatio-temporal features; the prediction steps of the improved neural network prediction model include:

[0064] S3.1: Parallelly allocate the input data to three hyperdimensional computing modules for high-dimensional feature encoding.

[0065] S3.2: Input the hyperdimensional feature vectors output by each module into the LSTM models of the corresponding channels for wave time series dynamic modeling.

[0066] S3.3: After the time series features output by the three LSTM channels are fused through the feature splicing layer, they are uniformly input into the CNN module for analyzing the wave spatial distribution characteristics.

[0067] S3.4: Input the multi-scale spatial features extracted by the CNN into the spatial-channel dual attention mechanism, strengthen the representation ability of key wave features through adaptive weight allocation, and finally output a high-precision wave situation prediction result.

[0068] Among them, the hyperdimensional computing technology enables the model to process higher-dimensional data, thereby capturing more potential information and features. The introduction of the Long Short-Term Memory network (LSTM) enables the model to learn long-term dependencies in the data and effectively cope with the complexity of time series data. The Convolutional Neural Network (CNN) is responsible for extracting local features and spatial information in the data, enhancing the model's representation ability. At the same time, the addition of the attention mechanism enables the model to automatically focus on the important parts of the input data, further improving the accuracy and robustness of the prediction. This combination method not only improves the prediction speed of the model but also significantly enhances the accuracy and reliability of the prediction results, providing a new and efficient solution for the prediction of time series data.

[0069] The present invention innovatively designs a brand-new attention mechanism, which mainly consists of two core modules: the spatial attention model and the channel attention model. In the model design, the output feature map of the CNN is equally divided into two parts, which are respectively fed into the spatial attention module and the channel attention module for parallel processing. The processing flow of the spatial attention module is as follows: First, global average pooling is performed on the input to capture spatial statistical information, then feature transformation is performed through a convolutional layer, followed by batch normalization operation, and finally, a spatial feature weight map is generated through the Sigmoid function. After global average pooling is used to compress the spatial dimension in the channel attention module, linear transformation of the channel dimension is achieved through a fully connected layer, the ReLU activation function is applied to enhance the non-linear expression ability, and finally, a channel feature weight map is generated through the Sigmoid function to complete the modeling of the dependencies between channels. These two modules each perform their own functions. Among them, the spatial attention model focuses on capturing the key information of the image or signal in the spatial dimension, while the channel attention model focuses on identifying and analyzing the important features in the channel (or time) dimension. Through the collaborative work of these two modules, the important parts in space and time (or channels) can be accurately extracted respectively, then these key information are weighted, and finally the two are organically combined, thereby significantly improving the accuracy and efficiency of information processing and recognition.

[0070] S4: Use the improved fireworks weighted average algorithm to perform locking control on the wave energy power generation device; the specific operation method is as follows:

[0071] S4.1: Use the improved fireworks weighted average algorithm to initialize N different positions, and obtain explosion fireworks and Gaussian fireworks by setting off fireworks at N different positions. Each set-off firework represents the optimal locking time under the locking control condition at the previous moment. The explosion fireworks and Gaussian fireworks represent the single optimal locking time and the global optimal locking time at the previous moment respectively.

[0072] S4.2: Set the maximum number of iterations MAX it, and start calculating the weighted average value. Rearrange all the locking times according to the two optimal locking times at the previous moment to obtain the number N of selected candidate solutions c , and calculate the weighted average position. The formula is as follows:

[0073]

[0074] In the formula, nP is the total number of the population, it is the current number of iterations. Then obtain the sum of the fitness values of all candidate solutions:

[0075]

[0076] In the formula, Sum Fitness is the sum of all fitness values, X i is the i-th candidate solution, Fitness is the function for calculating the fitness value. Then obtain the weighted average position X Miu :

[0077]

[0078] S4.3: During the iteration process, define the exploration and exploitation phases by the following formula:

[0079]

[0080] In this work, K1 is the conversion function, α is a random constant, rand is a random value from 0 to 1. Take the threshold of K1 as 0.5. When K1 ≥ 0.5, move according to the exploitation ability, and when K1 < 0.5, move according to the exploration ability.

[0081] S4.4: There are three strategies in total for the exploitation phase. The first one focuses on utilizing the search space, and this search space X i (it + 1) mainly depends on the current weighted average position and the personal best position X PersonBest obtained by the fireworks algorithm and the global optimal position X GlobalBest to determine, as follows:

[0082]

[0083] In the formula, w11, w12, and w13 are random values from 0 to 1, used to adjust the expansion of the search space.

[0084] The second strategy finally focuses on utilizing the search space established between the weighted average position of the whole population and the personal best position Under this strategy, the global optimal position is ignored. Compared with Strategy 1, this strategy reduces the search space and the search time. The formula is as follows:

[0085]

[0086] Wherein, w21 and w22 are random values between 0 and 1.

[0087] Strategy three focuses on the search space established between the weighted average position and the global optimal position, and the formula is as follows:

[0088]

[0089] Wherein, w31 and w32 also have a value range between 0 and 1. Compared with strategy one, strategy two and strategy three mainly improve the convergence speed and accuracy.

[0090] S4.5: When K1 < 0.5, the system is in the exploration stage, and it is mainly optimized through two exploration strategies. The formula of strategy one is as follows:

[0091]

[0092] S is the step size of Levy flight, which is a random walk pattern with heavy-tailed characteristics. Its step size follows the Levy distribution and is used in the optimization algorithm to enhance the global search ability; X GlobalBest,j (it) is the j-th position of the global optimal solution at the first iteration, X i,j (it + 1) is the j-th position of the i-th solution at the (it + 1)-th iteration.

[0093] In some cases, the global optimal value obtained by the algorithm in the area around the global optimal position is too different from the ideal value of the global optimal position. In this case, the weighted average algorithm will face the risk of converging to the local optimal. To solve this problem, the second strategy is adopted to adjust the search space, and the formula is as follows:

[0094] X i (it + 1) = rand * (UB min - LB min ) + LB min (9)

[0095] Wherein, LB min and UB min are the minimum values of the upper and lower bounds of the search space. Through this strategy, the search agent will move to a new position to search for a better area.

[0096] S4.6: Finally, check the boundary conditions and update the optimal position and the global optimal position of the single-body latching at the current time.

[0097] S4.7: After calculating the optimal latching position and time, maximize the wave through latching control.

Claims

1. A maximum power control method for wave energy based on hyperdimensional computing and deep learning, characterized in that: Including the steps: S1: Obtain the wave information of external waves through the sensors of the wave energy power generation device, including wave height and period; S2: Preprocess the obtained wave information and output the preprocessed wave information; S3: Input the wave information into the improved neural network prediction model for prediction to obtain future wave information; the improved neural network prediction model includes three layers of hyper-dimensional calculation models, three layers of LSTM models, one layer of CNN model, and one layer of attention mechanism model; This multi-modal fusion architecture realizes the hierarchical depth mining and collaborative optimization of wave spatio-temporal features; The prediction steps of the improved neural network prediction model include: S3.1: Parallelly distribute the input data to three hyper-dimensional calculation modules for high-dimensional feature encoding; S3.2: Respectively input the hyper-dimensional feature vectors output by each module into the LSTM models of the corresponding channels for wave time series dynamic modeling; S3.3: After the time series features output by the three LSTM channels are fused through the feature splicing layer, they are uniformly input into the CNN module for analyzing the wave spatial distribution characteristics; S3.4: Input the multi-scale spatial features extracted by the CNN into the spatial-channel dual attention mechanism, strengthen the representation ability of key wave features through adaptive weight allocation, and finally output a high-precision wave situation prediction result; S4: Use the improved fireworks weighted average algorithm to perform locking control on the wave energy power generation device.

2. The maximum power control method of wave energy based on hyperdimensional computing and deep learning according to claim 1, wherein: The attention mechanism model includes a spatial attention model and a channel attention model; the output feature map of the CNN model is equally divided into two parts, which are respectively fed into the spatial attention module and the channel attention module for parallel processing; The spatial attention module and the channel attention module operate in parallel. The spatial attention model focuses on capturing the key information in the spatial dimension of the image or signal, and the channel attention model focuses on identifying and analyzing the important features in the channel or time dimension.

3. The maximum power control method of wave energy based on hyperdimensional computing and deep learning according to claim 2, wherein: The processing flow of the spatial attention module is as follows: First, perform global average pooling on the input to capture spatial statistical information, then perform feature transformation through a convolutional layer, then perform batch normalization operation, and finally generate a spatial feature weight map through the Sigmoid function.

4. A maximum power control method for wave energy based on hyperdimensional computing and deep learning according to claim 2, characterized in that: The channel attention module adopts global average pooling to compress the spatial dimension, then realizes linear transformation of the channel dimension through a fully connected layer, applies the ReLU activation function to enhance the non-linear expression ability, and finally generates a channel feature weight map through the Sigmoid function to complete the modeling of the interdependence between channels.

5. A maximum power control method for wave energy based on hyperdimensional computing and deep learning according to claim 1, characterized in that: The specific operation method of step S4 is: S4.1: Use the improved fireworks weighted average algorithm to initialize N different positions, and obtain explosion fireworks and Gaussian fireworks by setting off fireworks at N different positions. Each set-off firework represents the optimal locking time in the previous moment under the implementation of locking control. The explosion fireworks and Gaussian fireworks represent the single optimal locking time and the global optimal locking time in the previous moment respectively; S4.2: Set the maximum number of iterations MAX it , and start calculating the weighted average. Rearrange all the blocking times according to the two optimal blocking times at the previous moment to obtain the number N of candidate solutions selected c , and calculate the position of the weighted average. The formula is as follows: In the formula, nP is the total number, it is the current loop count, and then obtain the sum of the fitness values of all candidate solutions: Where, Sum Fitness is the sum of all fitness values, X i is the i-th candidate solution, Fitness is the function for calculating fitness, and then the weighted average position X Miu is obtained as follows: S4.3: During the iteration process, the exploration and development stages are defined by the following formula: In this work, K1 is the conversion function, α is a random constant, rand is a random value between 0 and 1, and the threshold of K1 is taken as 0.

5. When K1 ≥ 0.5, it moves according to the development ability, while when K1 < 0.5, it moves according to the exploration ability; S4.4: There are three strategies in total during the development stage. The first one focuses on leveraging the search space, and the search space X i (it + 1) mainly depends on the current weighted average position and the personal best position X obtained by the fireworks algorithm PersonBest and the global optimal position X GlobalBest is determined as follows: In the formula, w11, w12, and w13 are random values between 0 and 1, which are used to adjust the expansion of the search space; The second strategy finally utilizes the search space established between the weighted average position of the entire population and the personal best position Under this strategy, the global optimal position is ignored. Compared with Strategy 1, this strategy reduces the search space and the search time. The formula is as follows: In the formula, w21 and w22 are random values between 0 and 1; Strategy three focuses on the search space established between the weighted average position and the global optimal position, and the formula is as follows: In the formula, w31 and w32 also have a value range between 0 and 1. Compared with strategy one, strategy two and strategy three mainly improve the convergence speed and accuracy; S4.5: When K1 < 0.5, the system is in the exploration stage, and it is mainly optimized through two exploration strategies. The formula for strategy one is as follows: S is the step size of Levy flight, which is a random walk pattern with heavy-tailed characteristics. Its step size follows the Levy distribution and is used in optimization algorithms to enhance the global search ability; X GlobalBest,j (it) is the j-th position of the global optimal solution at the first iteration, X i,j (it + 1) is the j-th position of the i-th solution at the (it + 1)-th iteration; In some cases, the global optimal value obtained by the algorithm is in the area around the global optimal position and is too different from the ideal value of the global optimal position. In this case, the weighted average algorithm will face the risk of converging to the local optimal. To solve this problem, the second strategy is adopted to adjust the search space, and the formula is as follows: X i (it + 1)= rand * (UB min - LB min )+ LB min (9) where LB min and UB min are the minimum values of the upper and lower bounds of the search space. Through this strategy, the search agent will move to a new position to search for a better area; S4.6: Finally, the boundary conditions are checked, and the optimal position and the global optimal position of the single-body latch at the current time are updated; S4.7: After calculating the optimal latch position and time, wave maximization is achieved through latch control.

Images