Offshore wind power prediction method and system

Abstract

The invention relates to the technical field of offshore wind power prediction, and discloses an offshore wind power prediction method and system. An anomaly detection threshold value is dynamically adjusted based on the wind speed change rate, abnormal value cleaning is carried out on historical data, and preprocessed data is obtained; establishing an extreme learning machine model, and initializing network structure parameters of the extreme learning machine model; an improved electric eel foraging optimization algorithm is adopted to optimize an input layer weight and a hidden layer threshold in the network structure parameters, and an optimized extreme learning machine model is generated; inputting the preprocessed data into the optimized extreme learning machine model for training, and constructing an offshore wind power prediction model; and predicting the real-time offshore wind power of the target port. The generalization performance of the model under the complex sea condition is effectively improved, a more accurate power prediction result is provided for power grid dispatching, and meanwhile the engineering timeliness requirement for real-time prediction of an offshore wind plant is met.

Description

Technical Field

[0001] This application relates to the technical field of offshore wind power prediction, and particularly to an offshore wind power prediction method and system. Background Art

[0002] With the transformation of the energy structure towards cleaner energy, offshore wind power has become an important direction for the development of renewable energy due to its resource endowment and location advantages. Compared with onshore wind power, the wind speed in the marine environment is higher and more stable, which can significantly improve the power generation efficiency. At the same time, the construction of offshore wind farms does not require the occupation of land resources, especially suitable for coastal economically developed areas with scarce land. In addition, as the population and industrial load centers, the coastal areas can effectively reduce the power transmission loss and improve the energy utilization efficiency by the near consumption of offshore wind power. However, the large-scale grid connection of offshore wind power also faces severe challenges: its output is affected by multiple factors such as marine meteorology, equipment status and grid connection conditions, with significant randomness and volatility. If the power prediction accuracy is insufficient, it will lead to a sharp increase in the grid frequency regulation pressure and the demand for reserve capacity, and even threaten the safe and stable operation of the power system. Therefore, high-precision offshore wind power prediction technology is the key foundation for realizing the efficient consumption of offshore wind power and the reliable regulation of the power grid.

[0003] The core goal of wind power prediction is to establish the mapping relationship between the output of the wind farm and the influencing factors by analyzing historical operation data and real-time environmental parameters, so as to achieve accurate estimation of future power generation. This technology has multiple values for the operation of the power system: at the grid dispatching level, accurate prediction results can optimize the formulation of power generation plans, reduce the frequency regulation burden of traditional power sources such as thermal power, and thus reduce the system operation cost; at the equipment operation and maintenance level, power prediction can be linked with the fan status monitoring system to identify potential faults through abnormal output and provide decision-making basis for preventive maintenance; at the power market level, the prediction accuracy directly affects the economic benefits of wind power participating in spot trading and ancillary services, and high-precision prediction helps to improve the market competitiveness of wind power.

[0004] Therefore, there is an urgent need for an offshore wind power prediction method that can efficiently and accurately predict the offshore wind power. Summary of the Invention

[0005] To solve the above technical problems, on the one hand, this application provides an offshore wind power prediction method, including the following steps: S1: Obtain the historical data of offshore wind power; S2: Dynamically adjust the outlier detection threshold based on the wind speed change rate, clean the outliers in the historical data, and obtain preprocessed data; S3: Establish an extreme learning machine model and initialize the network structure parameters of the extreme learning machine model; S4: Optimize the input layer weights and hidden layer thresholds in the network structure parameters by using the improved electric eel foraging optimization algorithm to generate an optimized extreme learning machine model; S5: Input the preprocessed data into the optimized extreme learning machine model for training to construct an offshore wind power prediction model; S6: Use the offshore wind power prediction model to predict the real-time offshore wind power of the target port.

[0006] Preferably, the calculation method for dynamically adjusting the anomaly detection threshold based on the wind speed change rate is as follows: ; In the formula, Uppert represents the upper limit for outlier judgment, Lowert represents the lower limit for outlier judgment, Q 75 represents the upper quartile, Q 25 represents the lower quartile, IQR t represents the interquartile range, represents the wind speed change rate, and λ represents the turbulence adjustment coefficient.

[0007] Preferably, the wind speed change rate is calculated using the following formula: ; In the formula, V t represents the wind speed value at time t, V t-1 represents the wind speed value at time t - 1, and V rated represents the rated wind speed value.

[0008] Preferably, the improvement of the improved electric eel foraging optimization algorithm is achieved through the following strategies: generating an initial population using the logistic chaotic map; introducing the golden sine strategy to determine the population movement path; and combining the adaptive random mutation strategy to adjust the individual positions of the newly generated individuals in each iteration.

[0009] Preferably, generating the initial population using the logistic chaotic map is specifically as follows: Randomly generate a vector, the dimension of the vector is Dim, and the value of the vector i in each dimension is a random number between 0 and 1; Use the logistic chaotic map formula to generate a sequence [s1,..., s i ,..., s N with the number of vectors being N, and the logistic chaotic map formula is as follows: ; In the formula, s i represents the i-th vector, s i-1 represents the (i - 1)-th vector, and f represents the control parameter of the chaotic map; According to the said sequence, distribution is carried out by the following formula to obtain the said initial population: ; In the formula, x i represents the position of the i-th electric eel corresponding to the i-th vector, Ub represents the upper limit of the value of the initial population, and Lb represents the lower limit of the value of the initial population.

[0010] Preferably, the golden sine strategy is introduced to determine the population movement path. Specifically, when the electric eel moves from the current position to the target position, the golden sine strategy is used to determine the population movement path to complete the update of the population position. The expression of the golden sine strategy is: ; In the formula, x i (t) represents the position of the i-th electric eel at time t, x i (t + 1) represents the position of the i-th electric eel at time t + 1, T1 is a random parameter with a value range of [0, 2π], T2 is a random parameter with a value range of [0, π], b1 and b2 respectively represent phase parameters, represents the golden ratio, X prey represents the global optimal position up to the current iteration.

[0011] Preferably, the position of the newly generated individual in each iteration is adjusted by combining the adaptive random mutation strategy. Specifically: ; In the formula, and represent the position of the newly generated individual after mutation and the original position before mutation in the n-th dimension, m represents the current iteration number, M represents the maximum iteration number, rand represents a random number between 0 and 1, represents the upper boundary of the solution in each iteration, represents the lower boundary of the solution in each iteration.

[0012] Preferably, the following formula is used to calculate the fitness value in the improved electric eel foraging optimization algorithm: ; In the formula, represents the fitness value, J represents the number of training samples, represents the true value of the wind power in the j-th training sample, represents the predicted value of the wind power in the j-th training sample.

[0013] Preferably, the historical data includes wind speed, wind direction and offshore power generation.

[0014] On the other hand, the present application also provides an offshore wind power prediction system for executing any of the above-mentioned offshore wind power prediction methods, including: a data module, a model building module and an output module; The data module is used to obtain historical data of offshore wind power; dynamically adjust the abnormal detection threshold based on the wind speed change rate, clean the historical data of abnormal values, and obtain pre-processed data; The model building module is connected to the data module and is used to establish an extreme learning machine model and initialize the network structure parameters of the extreme learning machine model; optimize the input layer weights and hidden layer thresholds in the network structure parameters using an improved electric eel foraging optimization algorithm to generate an optimized extreme learning machine model; input the preprocessed data into the optimized extreme learning machine model for training to build an offshore wind power prediction model; The output module is connected to the model building module and is used to predict the real-time offshore wind power of the target port by using the offshore wind power prediction model.

[0015] The embodiments of the present application have the following technical effects: The offshore wind power prediction method provided in this application effectively improves the accuracy and robustness of offshore wind power prediction by integrating dynamic threshold cleaning and intelligent optimization algorithm. First, the abnormal detection threshold is dynamically adjusted based on the wind speed change rate, which can adapt to the turbulence characteristics of the marine environment, accurately remove abnormal data while retaining the wind speed mutation characteristics, and solve the problem of excessive or insufficient cleaning of the traditional fixed threshold method under complex sea conditions, providing high-quality input for model training. Secondly, the improved electric eel foraging optimization algorithm is used to globally optimize the key parameters of the extreme learning machine, breaking through the limitation that traditional random initialization is prone to fall into local optimality, and by optimizing the parameter combination of the input layer weight and the hidden layer threshold, the model's ability to characterize the nonlinear power characteristics of offshore wind power is enhanced. Compared with conventional prediction methods, this technical solution, while maintaining the advantages of rapid training of extreme learning machines, effectively improves the generalization performance of the model under complex sea conditions through the collaborative innovation of data cleaning and parameter optimization, providing more accurate power prediction results for power grid dispatching, while meeting the engineering timeliness requirements of real-time prediction of offshore wind farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1It is a schematic diagram of the steps of an offshore wind power prediction method provided by an embodiment of the present application; Figure 2 It is a schematic diagram of the model framework of an offshore wind power prediction model provided by an embodiment of the present application; Figure 3 It is a schematic diagram of the process of an offshore wind power prediction method provided by an embodiment of the present application; Figure 4 It is a framework diagram for prediction by an offshore wind power prediction model provided by an embodiment of the present application; Figure 5 It is a schematic diagram of the prediction result of an offshore wind power prediction model provided by an embodiment of the present application; Figure 6 It is a comparison chart of the prediction result errors of different prediction models provided by an embodiment of the present application. Detailed implementation manners

[0018] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the scope protected by the present application.

[0019] It plays an important role in offshore wind power prediction. However, the particularity of offshore wind power makes traditional prediction methods face the problem of insufficient adaptability. First, the complexity of the marine environment far exceeds that of land. The coupling effect of meteorological and hydrological conditions such as typhoons, salt fog, and tides leads to stronger spatio-temporal volatility of wind speed and direction, and more significant non-linear characteristics of the power curve. Second, offshore wind farms are usually far from land, and sensor deployment and data transmission face high costs and high delays. Historical data may be subject to noise interference or local missing, which poses higher requirements for the robustness of the model. In addition, offshore wind turbines are in a high-humidity and high-salt corrosion environment for a long time, and the equipment aging rate is accelerated. Their power characteristics are prone to drift with the operating state, and the prediction model is required to have the ability of dynamic correction.

[0020] The current wind power prediction methods are mainly divided into three categories: physical models, statistical models, and machine learning. Physical models are based on fluid mechanics equations and the aerodynamic characteristics of wind turbines. Their advantage lies in the ability to analyze the physical process of wind speed propagation, but they rely on high-precision meteorological forecast data and wind turbine parameters. The complex meteorological conditions and dynamic changes in equipment parameters at sea often lead to the accumulation of model errors. Statistical models (such as ARIMA, Kalman filter, etc.) construct linear mapping relationships through time series analysis, but they have insufficient generalization ability when dealing with the non-linear relationship between wind speed and power, especially difficult to adapt to the multi-variable coupling characteristics of offshore wind power. Machine learning methods represented by neural networks have strong non-linear fitting ability, but their performance is limited by the quality of training data: on the one hand, the high noise and sparsity of offshore wind power data are prone to cause model overfitting; on the other hand, the structural parameters of traditional neural networks (such as weights and thresholds) are usually randomly initialized, and it is easy to fall into local optima when the data distribution is uneven, making it difficult to capture the dynamic change rules of offshore wind power. In addition, existing methods generally have the problem of high computational complexity, and offshore wind power prediction needs to take into account real-time requirements, which poses challenges to the lightweight design and optimization efficiency of the model.

[0021] In summary, existing prediction methods have obvious limitations in meeting the special requirements of offshore wind power. Physical models and statistical methods are difficult to handle high-dimensional non-linear relationships, while traditional machine learning methods are restricted by data quality and parameter sensitivity, resulting in insufficient prediction accuracy and stability. At the same time, the operation efficiency of complex models cannot meet the engineering requirements of real-time prediction in offshore wind farms. Therefore, there is an urgent need for an offshore wind power prediction method that takes into account accuracy, robustness, and computational efficiency to support its large-scale grid connection and the safe and economic operation of the power grid.

[0022] Based on this, the present application provides an offshore wind power prediction method, as Figure 1 shown, including the following steps: S1: Obtain historical data of offshore wind power; In some embodiments, the historical data includes wind speed, wind direction, and offshore power generation. Among them, wind speed and wind direction are used as the core input features, that is, the input parameters of the offshore wind power prediction model, and offshore power generation is used as the target variable of supervised learning, that is, the output parameter of the offshore wind power prediction model.

[0023] S2: Dynamically adjust the outlier detection threshold based on the wind speed change rate, clean the outlier in the historical data, and obtain preprocessed data; In some embodiments, the calculation method for dynamically adjusting the outlier detection threshold based on the wind speed change rate is: ; In the formula, Uppert represents the upper limit for outlier judgment, Lowert represents the lower limit for outlier judgment, Q75 represents the upper quartile, Q 25 represents the lower quartile, IQR t represents the interquartile range represents the wind speed change rate, and λ represents the turbulence adjustment coefficient.

[0024] In some embodiments, the wind speed change rate is calculated using the following formula: ; In the formula, V t represents the wind speed value at time t, V t-1 represents the wind speed value at time t - 1, V rated represents the rated wind speed value, that is, the wind speed when the wind turbine reaches the rated power output, which can be determined according to the technical specifications of the actual wind turbine model at the target port.

[0025] The above outlier cleaning is to perform preprocessing operations on all categories of data in the historical data respectively. The preprocessed historical wind speed data, historical wind direction data, and historical offshore power generation data are obtained.

[0026] The innovative point of the dynamic threshold calculation method is to introduce the wind speed change rate as an adjustment factor. The upper quartile Q 75 and the lower quartile Q 25 constitute the benchmark range of the data distribution, and the interquartile range IQRt reflects the data dispersion degree. The turbulence adjustment coefficient λ is set according to the sea area characteristics and is used to amplify or reduce the threshold adjustment range. When the wind speed change rate ΔVt increases, the threshold interval expands with the term (1 + λΔVt), which not only retains the real wind speed fluctuations in extreme weather such as typhoon passing through, but also effectively eliminates abnormal jump points. Compared with the fixed threshold method, this dynamic mechanism fully considers the characteristics of large-scale changes in offshore wind speed, which are prone to cause fluctuations in wind power, clearly reflects the influence mechanism of turbulence intensity in fluid mechanics on power fluctuations, solves the problem of over-cleaning or insufficient cleaning of the traditional fixed threshold method in complex sea conditions, provides high-quality input for model training, and provides a data basis for improving the accuracy of offshore wind power prediction.

[0027] Offshore wind farms are usually far from land, the cost of sensor deployment and real-time data collection is high, and historical data may be missing. In this application, the preprocessed historical data is used to replace the eliminated abnormal data and missing values through the median of the sliding window.

[0028] S3: Establish an extreme learning machine model and initialize the network structure parameters of the extreme learning machine model; Offshore wind power is affected by marine meteorological conditions (such as typhoons, salt fog, and tides) and wave dynamics. The spatio-temporal volatility of wind speed and direction is higher, resulting in more significant non-linear characteristics of power output. In this application, the Extreme Learning Machine (ELM), a new type of single-layer feedforward neural network designed for high-dimensional non-linear patterns, is selected to predict the offshore wind power. The network structure parameters of ELM are initialized, including the input layer weights, hidden layer thresholds, hidden layer nodes, activation functions, and so on. However, the input layer weights in ELM are randomly generated, usually random numbers from a uniform distribution or a normal distribution. The hidden layer threshold is the bias of each hidden layer neuron. In the activation function, this bias is added after the weighted sum of the input signals, and then processed by the activation function. For example, for a hidden layer node, its output may be the result of the activation function (such as sigmoid or ReLU) acting on the weighted sum of the input data plus the bias. The hidden layer threshold is also randomly set during initialization and remains unchanged during the training process. This random setting of the input layer weights and hidden layer thresholds makes the prediction result of the power unable to meet the accuracy requirements. Therefore, this application uses an improved electric eel foraging optimization algorithm to optimize the input layer weights and hidden layer thresholds in the network structure parameters, generating an optimized extreme learning machine model.

[0029] S4: Use an improved electric eel foraging optimization algorithm to optimize the input layer weights and hidden layer thresholds in the network structure parameters, generating an optimized extreme learning machine model; Due to the complexity of the offshore wind power environment, the wind farm is far from the land, the cost of deploying data acquisition equipment is high and data loss is likely to occur. When the input sample quantity is huge and the structure is complex, it causes the ELM calculation matrix to be ill-conditioned, and the network performance drops severely, thus affecting the prediction accuracy. Therefore, this paper uses an improved electric eel foraging optimization algorithm (IEEFO) to optimize the input layer weights and hidden layer thresholds to improve the prediction performance of ELM, and establishes an offshore wind power prediction model based on IEEFO-ELM. In this IEEFO-ELM offshore wind power prediction model, the input parameters of IEEFO are wind speed and wind direction, and the output parameters are the input layer weights and hidden layer thresholds of the optimized ELM. Based on these optimized input layer weights and hidden layer thresholds, ELM outputs the predicted offshore power generation. This method has a stronger global search ability and can better adapt to complex data structures and sample distributions.

[0030] In some embodiments, the improvement of the improved electric eel foraging optimization algorithm is achieved through the following strategy: S4A: Use logistic chaotic mapping to generate the initial population; In some embodiments, the use of logistic chaotic mapping to generate the initial population is specifically: Randomly generate a vector. The dimension of the vector is Dim, and the value of the vector i in each dimension is a random number between 0 and 1; Use the logistic chaotic mapping formula to generate a sequence [s1,…,s i ,…,s N , where N is also the number of initial solutions generated in the electric eel foraging optimization algorithm generated by the chaotic mapping. The logistic chaotic mapping formula is as follows: ; In the formula, s i represents the i-th vector, s i-1 represents the (i - 1)-th vector, and f represents the control parameter of the chaotic mapping; According to the sequence, perform the following distribution to obtain the initial population: ; In the formula, x i represents the position of the i-th electric eel corresponding to the i-th vector, that is, the initial solution corresponding to the electric eel. Ub represents the upper limit of the values of the initial population, and Lb represents the lower limit of the values of the initial population.

[0031] When implementing the Logistic chaotic mapping, the dimension Dim of s i is the number of variables. The two parameters of the extreme learning machine optimized by the electric eel foraging optimization algorithm are the input layer weight and the hidden layer threshold, so the value of Dim is 2. The dimension is an important parameter of the electric eel optimization algorithm. Each electric eel individual in the algorithm is equivalent to a possible solution in the solution space, and this solution contains all variables required to find the optimal value. During the generation process of the chaotic sequence, f is the control parameter of the chaotic mapping, and its value range is from 0 to 4. The larger the value of f, the more uniform the mapping distribution, and the best initialization effect of the electric eel foraging optimization algorithm. This method effectively utilizes the characteristics of the chaotic mapping to provide a more random and exploratory initial population for the EEFO algorithm, thereby improving the performance of the algorithm.

[0032] Historical data of offshore wind power may be unevenly distributed due to sensor failures or transmission delays. The initial population randomly initialized is likely to cause overfitting or underfitting of the prediction model. The initial population generated by the Logistic chaotic mapping in this application has ergodicity and uniformity, can cover a wider solution space, and avoid local convergence caused by data sparsity. It improves the adaptability of offshore wind power and enhances the robustness of the model to noise in scenarios with unstable data quality, provides a better search starting point for subsequent iterations, and improves the optimization effect of the input layer weight and the hidden layer threshold.

[0033] S4B: Introduce the golden sine strategy to determine the population movement path; In some embodiments, the introduction of the golden sine strategy determines the population movement path. Specifically, when the electric eel moves from the current position to the target position, the population movement path is determined by the golden sine strategy to complete the update of the population position. The expression of the golden sine strategy is as follows: ; In the formula, x i (t) represents the position of the i-th electric eel at time t, and x i (t + 1) represents the position of the i-th electric eel at time t + 1. T1 is a random parameter with a value range of [0, 2π], T2 is a random parameter with a value range of [0, π], and b1 and b2 respectively represent phase parameters. represents the golden ratio, and X prey represents the global optimal position up to the current iteration.

[0034] The traditional straight-line travel strategy is prone to missing potential optimal solutions in complex non-linear spaces, while the relationship between offshore wind power and wind speed is highly non-linear (such as the turbulence effect). In this application, the golden sine strategy adjusts the search step size and direction by introducing the golden ratio coefficient, enabling individual electric eels to expand the search range in a curved path and enhancing the global exploration ability. Specifically, the phase parameters b1 and b2 are calculated through the golden ratio τ = 0.618. This design makes the population movement path distributed in the unit circle according to the golden ratio, improving the diversity of search directions. T1 is a random parameter with a value range of [0, 2π], and T2 is a random parameter with a value range of [0, π], endowing the algorithm with random perturbation ability. The global optimal position X prey up to the current iteration serves as an attractor to guide the population movement direction. At the same time, the non-negativity of the movement step size is maintained through the absolute value operation of the sine function, avoiding ineffective oscillations. It enables individual electric eels to expand the search range in a curved path, being able to both deeply optimize locally and jump out of the current area to explore new solution spaces. For example, in the scenario of sudden wind speed change, the golden sine strategy captures the inflection point of the power drop through random phase perturbation, enhancing the adaptability of the model to extreme events. Further improving the optimization effect of the input layer weights and the hidden layer thresholds.

[0035] S4C: Combine the adaptive random mutation strategy to adjust the positions of the newly generated individuals in each iteration.

[0036] In some embodiments, the combination of the adaptive random mutation strategy to adjust the positions of the newly generated individuals in each iteration is specifically as follows: ; In the formula, and represent the positions of the newly generated individuals after mutation and the original positions before mutation in the n-th dimension, m represents the current iteration number, M represents the maximum iteration number, and rand represents a random number between 0 and 1. represents the upper bound of the solution in each iteration, represents the lower bound of the solution in each iteration. The upper and lower bounds are also determined according to the order of magnitude of the parameters to be optimized when being set.

[0037] The offshore wind speed is affected by multiple factors such as turbulence, tides, and waves, and the power data shows violent fluctuations and non-stationarity; the relationship between the wind direction and the power is complex and non-linear. The traditional Electric Eel Foraging Optimization (EEFO) algorithm relies on the current optimal solution as a "lighthouse", However, due to the strong non-linearity and high volatility of offshore wind power data, this traditional optimization algorithm may be difficult to adapt to the dynamic changes of this data, resulting in the model parameters being unable to effectively capture complex data patterns. And its strong dependence on the current optimal solution as a guide makes it easy to fall into a local optimum in the early stage of iteration, unable to further optimize the model parameters, thus limiting the prediction performance. Forcing individuals to gather towards the current optimal solution in each iteration causes the population diversity to decay rapidly, and the global search ability is limited. In a complex multi-modal solution space, the algorithm is prone to falling into a sub-optimal solution and difficult to jump out of the local optimum, such as the power inflection point corresponding to a sudden change in wind speed, and it cannot adapt to the characteristics of the drastic and unpredictable changes in environmental parameters such as offshore wind speed and wind direction. Therefore, the algorithm needs to have stronger dynamic adaptability, and this guiding mechanism based on the static optimal solution may not be able to respond to this change in time. Based on this, in each iteration of this application, some new individuals are generated based on the previous eel individuals, and the generation method is not limited, and any one of the conventional new individual generation methods can be used, such as crossover, mutation, etc.

[0038] After generating new individuals, a mutation operation is performed to randomly perturb these newly generated individuals to a certain extent to further increase the diversity and possibly jump out of the current local optimal region. The perturbation amplitudes of Gaussian mutation and Cauchy mutation are fixed and cannot be dynamically adjusted according to the iteration stage or environmental changes. However, in the offshore wind power scenario, a large range of exploration is required in the initial stage (such as a sudden drop in power caused by a typhoon), but the fixed amplitude may be insufficient; in the later stage, local optimization is required (such as adjusting parameters for a steady wind speed), but the fixed amplitude may be too large, reducing the convergence efficiency. The response to sudden wind speed changes (a sudden increase in ΔVt) is lagged, and the perturbation cannot match the actual power fluctuation amplitude; Gaussian mutation focuses on local fine search, and Cauchy mutation focuses on global jumps, but it is difficult to balance exploration and exploitation when using a single mutation strategy alone. For example, Gaussian mutation is prone to falling into a local optimum, such as a sub-optimal parameter combination under a steady wind speed; Cauchy mutation generates too many ineffective perturbations in the later stage, prolonging the convergence time.

[0039] In this application, the above-mentioned combined adaptive random mutation strategy is adopted to adjust the positions of newly generated individuals in each iteration. In order for the electric eel to obtain better positions, after each mutation, it is judged according to the fitness value of the electric eel population. Only when the fitness value of an individual is better will the optimal position be updated; otherwise, the original position will be retained.

[0040] (1) Realize dynamic disturbance amplitude adjustment Iterative stage adaptability: The disturbance amplitude coefficient (1 - 0.5×m / M) decreases linearly with the increase of the iteration number m, realizing "large-scale exploration in the initial stage and fine-tuning in the later stage". In the initial stage, the global search ability is enhanced (such as rapid response in typhoon scenarios), and in the later stage, excessive disturbance is avoided (such as efficient convergence in steady wind speeds).

[0041] Variable range adaptability: The disturbance amplitude is proportional to the upper and lower bounds of the variable (Ub−Lb), avoiding unreasonable disturbances caused by differences in different characteristic dimensions.

[0042] (2) Enhance diversity through random direction disturbance Generate a random direction disturbance between -1 and 1 through (−1 + 2×rand) to break the population convergence. In a complex solution space (such as multiple potential optimal parameters corresponding to multiple wind speed intervals), the random direction disturbance helps individuals explore different regions and reduces the risk of local optima.

[0043] (3) Adaptability to actual scenarios Response to wind speed mutation: When the wind speed change rate ΔVt is relatively high, the disturbance amplitude coefficient 0.1×(1 - 0.5×m / M) can still maintain a relatively large value (in the initial stage of iteration), ensuring sufficient disturbance to capture the sudden power drop.

[0044] Efficient convergence in the steady period: The disturbance amplitude decreases in the later stage of iteration, avoiding ineffective disturbances and improving the optimization efficiency of model parameters (for example, the RMSE of IEEFO-ELM in Table 1 is reduced by 41.9% compared with traditional ELM).

[0045] This formula is an adaptive mutation strategy, which is used to perform mutation operations on newly generated individuals. It is equivalent to extracting some solutions and updating the positions again. For example, extracting the positions of 10% of the individuals with the lowest fitness (the worst positions) at the current iteration number for update can improve the diversity of the population. It improves the global search ability of the algorithm and maintains the diversity of the population to prevent the electric eel foraging optimization algorithm from falling into local optimal solutions.

[0046] The improved electric eel foraging optimization algorithm IEEFO that optimizes the input layer weights and hidden layer thresholds of ELM uses the logistic chaotic map to generate the initial population and introduces the golden sine strategy to determine the population movement path; combines the adaptive random mutation strategy to adjust the positions of newly generated individuals in each iteration: The logistic chaotic map provides a high-quality initial solution. The golden sine strategy conducts global-local balance search based on this, while the adaptive mutation strategy performs a secondary perturbation on the preliminary optimization result, forming a closed-loop optimization process of "extensive coverage → in-depth development → dynamic correction", creating a linkage effect between initialization and search, and effectively improving the generalization performance of the model in the application scenarios of high volatility and strong nonlinearity in offshore wind power.

[0047] Dynamic adaptation to complex environments: The golden sine strategy responds to the change in wind speed direction through random phase, and the adaptive mutation strategy matches the wind speed fluctuation intensity through dynamic amplitude. The combination of the two ensures that the algorithm can flexibly adjust the search behavior under complex sea conditions and can provide more accurate power prediction results.

[0048] The chaotic map reduces invalid search paths, the golden sine strategy accelerates the positioning of high-quality regions, and the adaptive mutation strategy prevents local convergence. The three work together to significantly shorten the convergence time, taking into account both efficiency and accuracy, and meeting the engineering timeliness requirements of real-time prediction in offshore wind farms. As shown in Table 1, the RMSE of IEEFO-ELM is reduced by 41.9% compared with the traditional ELM, and at the same time, the generalization ability of the model is improved (R² = 0.9849).

[0049] In some embodiments, the fitness value in the improved electric eel foraging optimization algorithm is calculated using the following formula: ; In the formula, represents the fitness value, J represents the number of training samples, represents the true value of wind power in the j-th training sample, represents the predicted value of wind power in the j-th training sample.

[0050] This fitness value calculation function directly reflects the prediction accuracy of the model and guides the optimization algorithm to find the parameter combination that minimizes the training error. Exemplarily, in specific calculations, to avoid overfitting, k-fold cross-validation can also be used to calculate the fitness value. By minimizing the RMSE, the optimized extreme learning machine model shows stronger generalization ability on unknown data.

[0051] S5: Input the preprocessed data into the optimized extreme learning machine model for training to construct an offshore wind power prediction model, that is, the IEEFO-ELM model. The model framework of the offshore wind power prediction model constructed in this application is as Figure 2 shown, and the framework diagram for prediction through this model is as Figure 4 shown.

[0052] Exemplarily, the preprocessed data is divided into training data and test data, and the data allocation ratio is 9:1.

[0053] S6: Using the offshore wind power prediction model, predict the real-time offshore wind power of the target port. The specific prediction process can be seen in Figure 3 .

[0054] Adopt the wind power dataset of Tianjin Port Pacific International Container Terminal in November 2023, and use the offshore wind power prediction model constructed in this application to predict the wind power. In terms of parameter settings, the population size of the IEEFO algorithm is set to 50, and the maximum number of iterations is 400 to ensure that the algorithm can fully search for the optimal solution. The number of nodes in the hidden layer of the ELM model is set to 30 to construct a suitable model structure.

[0055] To comprehensively evaluate the performance of the IEEFO-ELM model in wind power prediction, this application compares it with three other prediction models. To ensure the stability and reliability of the evaluation results, 30 experiments are conducted for each prediction model. After the experiments are completed, the evaluation results of the four models are calculated and analyzed using the error index formula. Through a series of experiments and evaluations, we can more accurately understand the performance of various models in wind power prediction and provide strong support for practical applications. The prediction results of the IEEFO-ELM model for wind power are as Figure 5 shown, Figure 5 in which the predicted value curve obtained by the IEEFO-ELM model during wind power prediction is highly fitted with the actual value curve, fully demonstrating that the model has excellent performance and accuracy in the field of wind power prediction.

[0056] The comparison chart of the prediction result errors of different prediction models is as Figure 6 shown. In Figure 6 , the power fluctuation of the IEEFO-ELM model is represented by the blue line, the fluctuation of the EEFO-ELM model is represented by the green line, the fluctuation of the PSO-ELM model is represented by the pink line, and the fluctuation of the GWO-ELM model is represented by the cyan line. The abscissa represents 96 samples, each sample representing the wind power and its predicted value at a certain time period. Predict the wind power for 24 hours at 15-minute intervals, for a total of 24×4 = 96 sample points. It can be observed from the figure that the power fluctuation curve of the IEEFO-ELM model for wind power prediction is almost always below the curves of the EEFO-ELM, PSO-ELM, and GWO-ELM models, and the power fluctuations of the four models are also smaller than those of the traditional ELM prediction model. Compared with the PSO-ELM and GWO-ELM models, although the overall prediction effect of the EEFO-ELM is good, there are still several sample points with larger fluctuations compared to the IEEFO-ELM. To sum up, from the perspective of the wind power prediction fluctuation, among many prediction models, the IEEFO-ELM model performs better in wind power prediction.

[0057] This application also compares the prediction results of the above several prediction models from the dimensions of the mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination R 2 The comparison results are shown in Table 1.

[0058] Table 1 Comparison of Evaluation Indexes of PSO-ELM, GWO-ELM, EEFO-ELM, and IEEFO-ELM

[0059] To comprehensively evaluate the performance of the wind power prediction model, this application uses multiple error indexes, including the mean absolute percentage error, root mean square error, and coefficient of determination, etc., to ensure an accurate and comprehensive evaluation of the model's prediction ability.

[0060] Since there is no clear interval range for the two evaluation indexes of RMSE and MAPE when measuring prediction performance. To more comprehensively evaluate the prediction ability of the model, this application defines the coefficient of determination (R-Squared, R 2 ), where R 2 ∈[0,1], and its calculation formula is as follows: ; In the formula, K represents the total number of test samples, represents the actual value of the wind power in the test samples, represents the predicted value of the wind power in the test samples, and k represents the kth test sample.

[0061] According to the data in Table 1, it can be seen that the IEEFO-ELM model proposed in this chapter shows excellent performance in terms of prediction error. The MAPE index of the IEEFO-ELM model is 45.78, slightly larger than the other several models. Compared with EEFO-ELM, PSO-ELM, and GWO-ELM, the RMSE index of the IEEFO-ELM model reaches 7.163, which is significantly better than 10.50 of the EEFO-ELM model, 12.06 of the PSO-ELM model, 11.74 of the GWO-ELM model, and 12.3 of the original ELM model. Compared with the EEFO-ELM model, it is reduced by 31.8%, compared with the PSO-ELM model, it is reduced by 40.6%, compared with the GWO-ELM model, it is reduced by 38.9%, and compared with the traditional ELM model, it is reduced by 41.9%. And there is a significant improvement in both the power fluctuation amount and R 2 in two aspects. The R 2 index is 0.9849, and the fitting degree is greatly improved compared with other wind power prediction models, further confirming the effectiveness of this model in wind power prediction.

[0062] Exemplarily, the IEEFO-ELM model proposed in the present application can also be integrated into the port new energy management and control system. The IEEFO-ELM model proposed has better stability and higher prediction accuracy in wind power prediction, provides a more reliable method for the wind power prediction module of the port new energy management and control system, improves the system's optimization control ability for new energy, and improves the overall intelligent level of the port.

[0063] On the other hand, the present application also provides an offshore wind power prediction system that executes the offshore wind power prediction method described in any one of the above. The system includes: a data module, a model construction module, and an output module. The data module is used to obtain historical data of offshore wind power; dynamically adjust the anomaly detection threshold based on the wind speed change rate, clean the historical data of outliers, and obtain preprocessed data. The model construction module is connected to the data module and is used to establish an extreme learning machine model and initialize the network structure parameters of the extreme learning machine model; optimize the input layer weights and hidden layer thresholds in the network structure parameters by using an improved electric eel foraging optimization algorithm to generate an optimized extreme learning machine model; input the preprocessed data into the optimized extreme learning machine model for training to construct an offshore wind power prediction model. The output module is connected to the model construction module and is used to predict the real-time offshore wind power of the target port by using the offshore wind power prediction model.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting the power of offshore wind farms, characterized in that, It includes the following steps: S1: Obtain the historical data of offshore wind power; S2: Dynamically adjust the outlier detection threshold based on the wind speed change rate, clean the outliers in the historical data, and obtain preprocessed data; S3: Establish an extreme learning machine model and initialize the network structure parameters of the extreme learning machine model; S4: Optimize the input layer weights and hidden layer thresholds in the network structure parameters by using an improved electric eel foraging optimization algorithm to generate an optimized extreme learning machine model; S5: Input the preprocessed data into the optimized extreme learning machine model for training to construct an offshore wind power prediction model; S6: Use the offshore wind power prediction model to predict the real-time offshore wind power of the target port.

2. The offshore wind power prediction method according to claim 1, characterized in that The calculation method for dynamically adjusting the outlier detection threshold based on the wind speed change rate is as follows: ； Where Uppert represents the upper limit for outlier judgment, Lowert represents the lower limit for outlier judgment, Q 75 represents the upper quartile, Q 25 represents the lower quartile, IQR t represents the interquartile range, represents the wind speed change rate, and λ represents the turbulence adjustment coefficient.

3. The method for predicting the power of offshore wind turbines according to claim 2, wherein, The wind speed change rate is calculated by the following formula: ； Where, V t represents the wind speed value at time t, V t-1 represents the wind speed value at time t - 1, V rated represents the rated wind speed value.

4. A method for predicting the power of offshore wind turbines according to claim 1, characterized in that, The improvement of the improved electric eel foraging optimization algorithm is achieved through the following strategies: generating an initial population using logistic chaotic mapping; introducing the golden sine strategy to determine the population movement path; and combining the adaptive random mutation strategy to adjust the positions of newly generated individuals in each iteration.

5. The method for predicting the power of an offshore wind farm according to claim 4, wherein The specific method for generating the initial population using logistic chaotic mapping is as follows: Randomly generate a vector, the dimension of the vector is Dim, and the value of the vector i in each dimension is a random number between 0 and 1; Generate a sequence [s1, …, s i , …, s N with N vectors using the logistic chaotic mapping formula as follows: ； where s i represents the i-th vector, and s i-1 represents the (i - 1)-th vector, and f represents the control parameter of the chaotic map; According to the sequence, perform the following distribution to obtain the initial population: ； where x i represents the position of the i-th electric eel corresponding to the i-th vector, Ub represents the upper limit of the values of the initial population, and Lb represents the lower limit of the values of the initial population.

6. The offshore wind power prediction method according to claim 4, wherein The specific method for introducing the golden sine strategy to determine the population movement path is as follows: When the electric eel moves from the current position to the target position, determine the population movement path through the golden sine strategy to complete the update of the population position. The expression of the golden sine strategy is: ； where x i (t) represents the position of the i-th electric eel at time t, and x i (t + 1) represents the position of the i-th electric eel at time t + 1. T1 is a random parameter with a value range of [0, 2π], T2 is a random parameter with a value range of [0, π], and b1 and b2 represent phase parameters respectively, represents the golden ratio, and X prey represents the global optimal position up to the current iteration.

7. A method for predicting the power of an offshore wind farm according to claim 4, characterized in that The specific method for combining the adaptive random mutation strategy to adjust the positions of newly generated individuals in each iteration is as follows: ； In the formula, and represent the position after mutation and the original position before mutation of the newly generated individual in the nth dimension, m represents the current iteration number, M represents the maximum iteration number, rand represents a random number between 0 and 1, represents the upper bound of the solution in each iteration, represents the lower bound of the solution in each iteration.

8. A method for predicting the power of an offshore wind farm according to claim 1, characterized in that, The following formula is used to calculate the fitness value in the improved electric eel foraging optimization algorithm: ； In the formula, represents the fitness value, J represents the number of training samples, represents the true value of wind power in the j-th training sample, represents the predicted value of wind power in the j-th training sample.

9. A method for predicting the power of an offshore wind farm according to claim 1, characterized in that, The historical data includes wind speed, wind direction, and offshore power generation.

10. An offshore wind power prediction system, characterized in that, Implementing an offshore wind power prediction method according to any one of claims 1-9 includes: a data module, a model construction module, and an output module; The data module is used to obtain the historical data of offshore wind power; dynamically adjust the outlier detection threshold based on the wind speed change rate, clean the outliers in the historical data, and obtain preprocessed data; The model construction module is connected to the data module and is used to establish an extreme learning machine model, initialize the network structure parameters of the extreme learning machine model; optimize the input layer weights and hidden layer thresholds in the network structure parameters by using an improved electric eel foraging optimization algorithm to generate an optimized extreme learning machine model; input the preprocessed data into the optimized extreme learning machine model for training to construct an offshore wind power prediction model; The output module is connected to the model construction module and is used to use the offshore wind power prediction model to predict the real-time offshore wind power of the target port.

Images