Method, device, electronic equipment and storage medium for predicting wind power

Abstract

This specification provides a method, apparatus, electronic device, and storage medium for predicting wind power. The method comprises two stages: similar wind turbine screening and power prediction. In the similar wind turbine screening stage, addressing the problem that existing similar wind turbine screening methods rely on power sequence observation characteristics and struggle to reflect the power generation patterns of wind turbines, a Gaussian Hidden Markov Model is used to evaluate the similarity between wind turbines at the power sequence generation mechanism level, screening out wind turbines with similar power generation patterns to the target wind turbine. In the power prediction stage, addressing the problem that existing hedging LSTM networks, under a multi-layered modeling mechanism with the same target, have a strong dependence on a few dominant layers, a residual hedging LSTM network is proposed: the output of the first-layer LSTM serves as the baseline prediction value, used to fit the true value; the second to Lth layers of LSTM are used to fit the residual between the true value and the baseline prediction value, thereby reconstructing the prediction target for each layer. Through a collaborative mechanism between the baseline prediction and residual correction, the accuracy of wind power prediction is improved.

Description

Technical Field

[0001] This specification relates to the field of wind power prediction technology, and more particularly to a wind power prediction method, apparatus, electronic device and storage medium. Background Technology

[0002] Because wind speed is affected by factors such as meteorological conditions, atmospheric boundary layer movement, and terrain, the output power of wind turbines exhibits significant intermittency and fluctuation. Therefore, accurate prediction of future wind turbine power is of great significance for ensuring the safe and stable operation of the power system, improving the efficiency of electricity market transactions, and optimizing the operation and maintenance of wind farms.

[0003] In existing wind power prediction technologies, a common approach is to construct a time-series prediction model based on the historical power data of the target wind turbine itself. Typical examples include power prediction models based on LSTM (Long Short-Term Memory) and its variants. Some schemes further incorporate historical data of wind turbines in the same wind farm and with a certain degree of similarity to the target wind turbine as auxiliary information during the model input construction stage. Therefore, it is necessary to first screen similar wind turbines through wind turbine power sequence similarity assessment methods.

[0004] However, existing wind power prediction technologies still have shortcomings in two aspects: screening for similar wind turbines and designing prediction models. On the one hand, existing wind turbine power sequence similarity assessment methods mainly rely on the observation characteristics of historical power sequences to evaluate the similarity between wind turbines, making it difficult to reflect the inherent laws of wind turbine power generation. On the other hand, existing LSTM-based hedging LSTM networks suffer from a strong dependence on a few dominant layers under the multi-layer same-objective modeling mechanism. Summary of the Invention

[0005] In view of the above, one or more embodiments of this specification provide the following technical solutions: According to a first aspect of one or more embodiments of this specification, a method for predicting wind power is proposed, the method comprising: Obtain the historical power sequence of the target wind turbine and other wind turbines in the same wind farm; Based on the Gaussian Hidden Markov Model, the power sequence of the target wind turbine and each candidate wind turbine selected from the other wind turbines is probabilistically modeled, and the similarity of the generation mechanism between the target wind turbine and each candidate wind turbine is evaluated according to the similarity between the models, so as to select a number of similar wind turbines with similar generation mechanism to the target wind turbine from the candidate wind turbines. The historical power data of the target wind turbine and the similar wind turbines are input into the power prediction model to obtain the future power prediction value of the target wind turbine; The power prediction model includes an L-layer LSTM, wherein the first layer LSTM serves as the baseline prediction branch, which is used to fit the future actual power of the target wind turbine and output the baseline prediction value; the second to Lth layers LSTM serve as the residual prediction branch, which are used to fit the residual between the future actual power of the target wind turbine and the baseline prediction value and output the residual prediction value; the future power prediction value is calculated based on the baseline prediction value and the residual prediction value.

[0006] Optionally, the step of probabilistically modeling the power sequences of the target wind turbine and each candidate wind turbine selected from the other wind turbines based on a Gaussian Hidden Markov Model, and evaluating the similarity of the generation mechanisms of the target wind turbine and each candidate wind turbine based on the similarity between the models, so as to select a number of similar wind turbines from the candidate wind turbines that have similar generation mechanisms to the target wind turbine, includes: Calculate the mutual information between the historical power sequence of the target wind turbine and the historical power sequences of other wind turbines in the same wind farm, and select candidate wind turbines from the other wind turbines based on the estimated mutual information value; The optimal number of hidden states in the Gaussian Hidden Markov Model of the target wind turbine is determined, and the optimal number of hidden states is used as the number of hidden states in the Gaussian Hidden Markov Model of each candidate wind turbine, so as to construct the Gaussian Hidden Markov Model of the target wind turbine and each candidate wind turbine respectively. The generation mechanism similarity index between each candidate wind turbine and the target wind turbine is calculated based on the constructed Gaussian Hidden Markov Model, and the similar wind turbines are selected in descending order of generation mechanism similarity index.

[0007] Optionally, constructing Gaussian Hidden Markov Models for the target wind turbine and each candidate wind turbine includes: For the target wind turbine, the Gaussian Hidden Markov Model with estimated parameters corresponding to the optimal number of hidden states is selected as the Gaussian Hidden Markov Model of the target wind turbine. For any of the candidate wind turbines, the model parameters are estimated using the Baum-Welch algorithm based on the optimal number of hidden states to construct a Gaussian hidden Markov model for the candidate wind turbine.

[0008] Optionally, the step of calculating the generation mechanism similarity index between each candidate wind turbine and the target wind turbine based on the constructed Gaussian Hidden Markov Model includes: The power sequence of the candidate wind turbines is mapped to the normalized space of the target wind turbine, and the first average log-likelihood is calculated under the Gaussian hidden Markov model of the target wind turbine. The power sequence of the target wind turbine is mapped to the normalized space of the candidate wind turbine, and the second average log-likelihood is calculated under the Gaussian hidden Markov model of the candidate wind turbine. The mean of the first average log-likelihood and the second average log-likelihood is calculated and used as an index of the generation mechanism similarity between the candidate wind turbine and the target wind turbine.

[0009] Optionally, the future power prediction value is obtained by weighting and adding the baseline prediction value to the residual prediction values ​​of the LSTM outputs of each layer from the second to the Lth layer.

[0010] Optionally, the training process of the power prediction model includes: The baseline prediction loss is calculated using the actual power value of the target wind turbine as the supervision signal of the first layer LSTM. The difference between the actual power value of the target wind turbine and the baseline prediction value output by the first layer LSTM is used as the supervision signal of the second to Lth layers LSTM, and the residual prediction loss of the second to Lth layers LSTM is calculated. The weights corresponding to the residual prediction branches from the second to the Lth layer are dynamically updated based on the residual prediction loss, so that the branches with smaller residual prediction losses receive larger weight values. Calculate the weighted sum of the residual prediction values ​​and corresponding weight values ​​of the LSTM output from the second to the Lth layer, and add it to the baseline prediction value to obtain the power prediction result; The parameters of each LSTM layer of the power prediction model are updated based on the baseline prediction loss and each residual prediction loss, wherein the parameters of the first LSTM layer are updated based on the baseline prediction loss and each residual prediction loss; The parameters of the LSTM layer are based on the first layer. Layer to the first The residual prediction loss corresponding to the layer is updated, where, The value of is greater than or equal to 2 and less than or equal to L.

[0011] Optionally, dynamically updating the weights corresponding to the residual prediction branches from the second to the Lth layer based on the residual prediction loss includes: Calculate the prediction error of the residual prediction value for each residual prediction branch from the second layer to the Lth layer respectively; The weights corresponding to each residual branch are multiplied by a decay factor determined based on the prediction error of each residual branch to obtain updated weight values, wherein the decay factor is an exponential decay function of the prediction error.

[0012] According to a second aspect of one or more embodiments of this specification, a wind power prediction device is provided, the device comprising: The data acquisition unit acquires the historical power sequence of the target wind turbine and other wind turbines in the same wind farm. The similar wind turbine screening unit performs probabilistic modeling of the power sequences of the target wind turbine and each candidate wind turbine selected from the other wind turbines based on the Gaussian Hidden Markov Model, and evaluates the similarity of the generation mechanism between the power sequences based on the similarity between the models, so as to screen out a number of similar wind turbines with similar generation mechanisms to the target wind turbine from the candidate wind turbines. The power prediction unit inputs the historical power data of the target wind turbine and the similar wind turbines into the power prediction model to obtain the future power prediction value of the target wind turbine. The power prediction model includes an L-layer LSTM, wherein the first layer LSTM serves as the baseline prediction branch, used to fit the future actual power of the target wind turbine and output the baseline prediction value; the second to Lth layers of LSTM are used to fit the residual between the future actual power of the target wind turbine and the baseline prediction value, and output the residual prediction value; the future power prediction value is calculated based on the baseline prediction value and the residual prediction value.

[0013] According to a third aspect of one or more embodiments of this specification, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the aforementioned method by executing the executable instructions.

[0014] According to a fourth aspect of one or more embodiments of this specification, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the aforementioned method.

[0015] According to a fifth aspect of one or more embodiments of this specification, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the aforementioned method.

[0016] As can be seen from the above embodiments, the wind power prediction method provided in this specification includes two stages: similar wind turbine screening and power prediction. In the similar wind turbine screening stage, addressing the problem that existing similar wind turbine screening methods rely on power sequence observation characteristics and are difficult to reflect the power generation law of wind turbines, a Gaussian Hidden Markov Model is used to evaluate the similarity between wind turbines at the power sequence generation mechanism level, screening out wind turbines with similar power generation laws to the target wind turbine. In the power prediction stage, addressing the problem that existing hedging LSTM networks have a strong dependence on a few dominant layers under multi-layer same-target modeling mechanisms, a residual hedging LSTM network is proposed. This model uses the output of the first layer LSTM as the baseline prediction value to fit the true value; the second to Lth layers of LSTM are used to fit the residual between the true value and the baseline prediction value, thereby reconstructing the prediction target of each layer. Through the synergistic mechanism of baseline prediction and residual correction, the accuracy of wind power prediction is improved. Attached Figure Description

[0017] Figure 1 This is a flowchart of a wind power prediction method provided in an exemplary embodiment.

[0018] Figure 2 This is a flowchart of a screening method for similar wind turbines provided in an exemplary embodiment.

[0019] Figure 3 This is a schematic diagram of the structure of a power prediction model provided in an exemplary embodiment.

[0020] Figure 4 This is a schematic diagram of the structure of a device provided in an exemplary embodiment.

[0021] Figure 5 This is a block diagram of a wind power prediction device provided in an exemplary embodiment. Detailed Implementation

[0022] Because wind speed is affected by factors such as meteorological conditions, atmospheric boundary layer movement, and terrain, the output power of wind turbines exhibits significant intermittency and fluctuation. Therefore, accurate prediction of future wind turbine power is of great significance for ensuring the safe and stable operation of the power system, improving the efficiency of electricity market transactions, and optimizing the operation and maintenance of wind farms.

[0023] In existing wind power prediction technologies, a common approach is to construct a time-series prediction model based on the historical power data of the target wind turbine itself. Typical examples include power prediction models based on LSTM (Long Short-Term Memory) and its variants. Some schemes further incorporate historical data of wind turbines in the same wind farm and with a certain degree of similarity to the target wind turbine as auxiliary information during the model input construction stage. Therefore, it is necessary to first screen similar wind turbines through wind turbine power sequence similarity assessment methods.

[0024] However, existing wind power prediction technologies still have shortcomings in two aspects: screening for similar wind turbines and designing prediction models. On the one hand, existing wind turbine power sequence similarity assessment methods mainly rely on the observation characteristics of historical power sequences to evaluate the similarity between wind turbines, making it difficult to reflect the inherent laws of wind turbine power generation. On the other hand, existing LSTM-based hedging LSTM networks suffer from a strong dependence on a few dominant layers under the multi-layer same-objective modeling mechanism.

[0025] This specification provides a wind power prediction scheme. It probabilistically models the historical power sequences of the target and candidate wind turbines within a wind farm, assessing the similarity between turbines at the level of their power sequence generation mechanisms. This allows for the selection of several similar wind turbines from the candidate turbines whose generation mechanisms are similar to the target turbine. The historical power data of both the target and similar turbines are then input into a power prediction model to predict the future power of the target turbine. The power prediction model is based on a multi-layer LSTM network. The first LSTM layer outputs a baseline prediction value for the target turbine, while the remaining LSTM layers use the residuals relative to the baseline prediction value as monitoring signals. This forms a prediction mechanism that combines baseline prediction with residual correction to improve the accuracy of wind power prediction.

[0026] The wind power prediction scheme provided in this specification can be executed by a server, by a terminal device in conjunction with a server, or by a distributed computing cluster or cloud service platform. The server is a physical server containing an independent host, or a virtual server hosted by a host cluster. The terminal device can be a personal computer, mobile phone, tablet, laptop, workstation, industrial control computer, etc., and one or more embodiments in this specification do not limit this.

[0027] In practical deployment scenarios, an architecture combining offline training and online prediction is typically adopted. The training phase demands significant computing resources, requiring extensive iterative computations and parameter updates. High-performance computing clusters or graphics processing unit (GPU) clusters can be used to train the power prediction model. After training, the model parameters are exported and deployed to the prediction endpoint. The prediction endpoint can be a cloud-based inference server that receives real-time power data collected from the wind farm and returns prediction results; it can also be a locally deployed edge computing node or station control-level server at the wind farm, enabling low-latency real-time response; or it can be a terminal device (such as a wind farm monitoring workstation, mobile operation and maintenance terminal, etc.) directly used by operation and management personnel. Those skilled in the art can flexibly choose the model training and deployment method based on the computing resources, real-time requirements, and data transmission costs of the actual application scenario.

[0028] Figure 1 This is a flowchart of a wind power prediction method provided in an exemplary embodiment.

[0029] Please refer to Figure 1 The wind power prediction method may include the following steps: Step 102: Obtain the historical power sequence of the target wind turbine and other wind turbines in the same wind farm.

[0030] In some embodiments, each wind turbine in a wind farm is equipped with a monitoring and data acquisition system that can record the average output power of the turbine within a corresponding sampling time period at fixed sampling intervals (e.g., every 10 minutes or every 15 minutes). By arranging the power data (such as active power) of the same wind turbine in chronological order, a historical power sequence for that wind turbine can be formed.

[0031] Step 104: Based on the Gaussian Hidden Markov Model, the power sequence of the target wind turbine and each candidate wind turbine selected from the other wind turbines is probabilistically modeled, and the similarity of the generation mechanism between the target wind turbine and each candidate wind turbine is evaluated according to the similarity between the models, so as to select a number of similar wind turbines with similar generation mechanisms to the target wind turbine from the candidate wind turbines.

[0032] In some embodiments, a Gaussian Hidden Markov Model (GHMM) can model a power sequence as a stochastic process governed by a finite number of unobservable hidden states. The transition patterns between these hidden states and the observation probability distributions corresponding to each hidden state jointly characterize the power sequence generation mechanism. Therefore, if the GHMMs corresponding to two wind turbines have high similarity, they can be considered to have similar power generation mechanisms.

[0033] For example, for the target wind turbine and each candidate wind turbine, corresponding Gaussian Hidden Markov Models can be constructed to describe the probabilistic generation process of their respective power sequences. Then, based on the Gaussian Hidden Markov Models corresponding to different wind turbines, a generation mechanism similarity index between each candidate wind turbine and the target wind turbine is calculated. The larger the generation mechanism similarity index, the more similar the corresponding Gaussian Hidden Markov Models are, and thus the more similar the power generation mechanisms of the candidate wind turbine and the target wind turbine are. Next, several wind turbines with the highest rankings are selected as similar wind turbines according to the index from largest to smallest. The construction method of the Gaussian Hidden Markov Model, the determination method of the number of hidden states, and the calculation method of the generation mechanism similarity index will be described in detail in subsequent embodiments.

[0034] Step 106: Input the historical power data of the target wind turbine and similar wind turbines into the power prediction model to obtain the future power prediction value of the target wind turbine. The power prediction model includes an L-layer LSTM, wherein the first layer LSTM serves as the baseline prediction branch, used to fit the actual future power of the target wind turbine and output the baseline prediction value; the second to Lth layers of LSTM serve as the residual prediction branch, used to fit the residual between the actual future power of the target wind turbine and the baseline prediction value, and output the residual prediction value; the future power prediction value is calculated based on the baseline prediction value and the residual prediction value.

[0035] In some embodiments, the historical power data of the target wind turbine, as well as the historical power data of similar wind turbines selected from the pool, can be input into a pre-trained power prediction model for inference.

[0036] The power prediction model includes an L-layer LSTM, where L is greater than 1, and each LSTM layer transmits temporal features layer by layer in the time dimension.

[0037] In the specific structure of the power prediction model, the first-layer LSTM serves as the baseline prediction branch, outputting a baseline power prediction value. Specifically, the first-layer LSTM receives a multivariate time-series input composed of historical power data from the target wind turbine and similar wind turbines, and extracts time-series features from this input sequence. The hidden state generated at the last time step contains the main information for predicting the future power of the target wind turbine. This hidden state is then input into a linear mapping layer, where a linear transformation maps the hidden state to a scalar value. This scalar value is the baseline power prediction value of the target wind turbine predicted by the first-layer LSTM, reflecting the power prediction model's main estimate of the target wind turbine's future power.

[0038] The second to Lth layers of the LSTM serve as residual prediction branches to predict the residuals relative to the baseline power prediction. Specifically, the second to Lth layers of the LSTM receive the hidden states from the output of the previous layer of the LSTM, extract the final hidden state of that layer at the last time step, and input it into their respective linear mapping layers to obtain the corresponding residual estimates. Unlike the first layer of the LSTM, the supervision signal of the second to Lth layers of the LSTM is not the actual power value of the target wind turbine itself, but rather the residual between the actual power value and the baseline power prediction. In this way, the first layer of the LSTM undertakes the main prediction task, while the second to Lth layers of the LSTM focus on learning the error components remaining after the main prediction.

[0039] Based on this, the power prediction model weights and combines the baseline power prediction value with the residual prediction values ​​corresponding to the second to Lth layers of the LSTM according to the weights corresponding to each residual prediction value (i.e., the weights corresponding to each residual prediction branch) to obtain the future power prediction value of the target wind turbine.

[0040] Therefore, the wind power prediction method provided in this embodiment includes two stages: similar wind turbine screening and power prediction. In the similar wind turbine screening stage, addressing the problem that existing similar wind turbine screening methods rely on power sequence observation characteristics and struggle to reflect the power generation patterns of wind turbines, a Gaussian Hidden Markov Model is used to evaluate the similarity between wind turbines at the power sequence generation mechanism level, screening out wind turbines with similar power generation patterns to the target wind turbine. In the power prediction stage, addressing the problem that existing hedging LSTM networks have a strong dependence on a few dominant layers under multi-layer same-target modeling mechanisms, a residual hedging LSTM network is proposed. This model uses the output of the first-layer LSTM as the baseline prediction value to fit the true value; the second to Lth layers of LSTM are used to fit the residual between the true value and the baseline prediction value, thereby reconstructing the prediction target of each layer. Through the synergistic mechanism of baseline prediction and residual correction, the accuracy of wind power prediction is improved.

[0041] In some embodiments, after obtaining the historical power sequence of the target wind turbine and other wind turbines in the same wind farm, it is necessary to select several similar wind turbines that are similar to the target wind turbine in terms of power generation mechanism from the other wind turbines, and use them as reference inputs for subsequent power prediction models.

[0042] Please refer to Figure 2 The screening process for similar wind turbines may include the following steps: Step 202: Calculate the mutual information between the historical power sequence of the target wind turbine and the historical power sequences of other wind turbines in the same wind farm, and select several other wind turbines as candidate wind turbines based on the mutual information estimate.

[0043] In some embodiments, since the optimal number of hidden states in the Gaussian Hidden Markov Model (HMM) may differ for different wind turbines, if the complete optimization of the hidden state number and model construction are performed independently for each wind turbine in the wind farm, multiple Gaussian HMMs with different hidden state numbers need to be trained repeatedly, and the optimal model needs to be determined based on the model selection criteria. This results in significant computational overhead and is difficult to meet computational efficiency requirements in large-scale wind farm scenarios. To improve the efficiency of similar wind turbine screening, it is advisable to first screen candidate wind turbines when screening similar wind turbines.

[0044] In some embodiments, mutual information is a fundamental concept in information theory used to measure the degree of statistical dependence between two random variables; a larger value generally indicates a stronger dependency between the two variables. Specifically, the mutual information between the target wind turbine and any other wind turbine in the wind farm can be estimated using a nonparametric mutual information estimation method based on nearest neighbor search. This method does not require prior assumptions that the data follows a specific probability distribution and can adapt to the non-Gaussian and non-stationary characteristics of wind power sequences. Subsequently, all wind turbines in the wind farm, excluding the target wind turbine, are sorted in descending order of their estimated mutual information values, and the top-ranked turbines are selected as candidate turbines. Through this pre-screening, only Gaussian Hidden Markov Models need to be built for the candidate turbines, rather than modeling each turbine in the wind farm individually, thus significantly reducing computational overhead.

[0045] Assuming, This represents the total number of wind turbines in the wind farm. The total number of time steps. Given the set of indices for all wind turbines, retrieve the index of the target wind turbine. The historical power data of all wind turbines in the same wind farm form a matrix. for: in, Represents wind turbine The power sequence, The value ranges from 1 to .

[0046] Next, a nonparametric mutual information estimation method based on nearest neighbor search can be used to calculate the target wind turbine. Compared with other fans Estimated value of mutual information between : in, This represents the mutual information estimation function.

[0047] Then, sorting the mutual information estimates from largest to smallest, the top [values] were selected. Each wind turbine index constitutes a candidate wind turbine set. : Step 204: Determine the optimal number of hidden states in the Gaussian Hidden Markov Model of the target wind turbine, and use the optimal number of hidden states as the number of hidden states in the Gaussian Hidden Markov Model of each candidate wind turbine, so as to construct the Gaussian Hidden Markov Model of the target wind turbine and each candidate wind turbine respectively.

[0048] A Gaussian Hidden Markov Model (GHMM) is a doubly stochastic process model consisting of an unobservable finite-state Markov chain and the Gaussian observation probability distributions corresponding to each hidden state. In wind power prediction scenarios, each hidden state can be understood as a statistically significant potential operating mode of the wind turbine at a certain moment, such as high-power stability, low-power fluctuation, rapid power increase, or rapid power decrease. The state transition probabilities between hidden states can characterize the dynamic evolution of the power sequence between different operating modes, while the Gaussian observation distributions corresponding to the hidden states describe the statistical characteristics of the power values ​​under specific operating modes. By jointly modeling the evolution of hidden states and the characteristics of observation distributions, GHMM can characterize the power sequence at the generation mechanism level, thus providing a basis for assessing the similarity of the generation mechanisms of power sequences from different wind turbines.

[0049] In some embodiments, the optimal number of hidden states in the Gaussian Hidden Markov Model of the target wind turbine can be determined by the following steps: First, the power sequence of the target wind turbine is divided into a training set and a validation set in chronological order. The division ratio can be set according to the actual amount of data. For example, the first 80% of the sequence can be used as the training set and the last 20% as the validation set.

[0050] Assuming, using Representing the proportion of the training set, the power sequence of the target wind turbine. Divided into training sets according to time sequence and verification set : Next, the complete power sequence of the target wind turbine is standardized based on the mean and standard deviation of the training set to reduce the impact of the power numerical scale on the model parameter estimation.

[0051] Assume that the training set The mean is expressed as The standard deviation is expressed as : Define the normalized operator ,Will pass Transformed into a normalized power sequence : Next, a candidate set of hidden states is constructed, for example, consisting of positive integers from 2 to 10. For each candidate hidden state in this set, the parameters of the corresponding Gaussian Hidden Markov Model are estimated on the training set using the Baum-Welch algorithm. The Baum-Welch algorithm is a parameter estimation algorithm for Hidden Markov Models based on the Expectation-Maximization (EM) concept. It uses a forward-backward algorithm to calculate the posterior probability of the hidden state and the posterior probability of the hidden state transition at adjacent time steps, and iteratively updates the model parameters between the expectation step and the maximization step until the algorithm converges, thereby obtaining the model parameters that make the log-likelihood function of the observed data locally optimal given the number of hidden states. Subsequently, the log-likelihood value of the trained Gaussian Hidden Markov Model is calculated based on the observation sequence of the validation set to evaluate the model's ability to interpret the generated probabilities of data not involved in the training.

[0052] Optionally, to determine the optimal number of hidden states in the Gaussian Hidden Markov Model of the target wind turbine, the Bayesian Information Criterion (BIC) can be introduced as a model selection criterion. For each candidate number of hidden states, the BIC value is calculated based on the log-likelihood value of the validation set and the number of free parameters of the model at that number of hidden states. The candidate number of hidden states with the smallest BIC value is selected as the optimal number of hidden states for the target wind turbine.

[0053] Assume that the normalized power sequence The number of hidden states is The Gaussian Hidden Markov Model (GHMM) is adopted. Representative model in The hidden state at time t, then the parameter set of the model and The range of values ​​for is: in, Let be the initial state probability vector, whose elements are... The definition of is: The matrix represents the state transition probability matrix, and its elements are... The definition of is: in, Hidden state The standardized mean and covariance of the corresponding observed distribution, i.e.: use The set representing the number of candidate hidden states, for any In the training set Estimating model parameters using the Baum-Welch algorithm And calculate the validation set. Log-likelihood: have Number of free parameters of a hidden state GHMM for: Next, the number of candidate hidden states is calculated. BIC value of the validation set : in, The number of hidden states corresponding to the minimum BIC in the validation set is selected as the optimal number of hidden states for the target wind turbine GHMM. : In some embodiments, after determining the optimal number of hidden states of the target wind turbine, the optimal number of hidden states is used as the number of hidden states of the Gaussian Hidden Markov Model of each candidate wind turbine, and the Gaussian Hidden Markov Models of the target wind turbine and each candidate wind turbine are constructed respectively.

[0054] Specifically, for the target wind turbine, the Gaussian Hidden Markov Model with estimated parameters corresponding to the optimal number of hidden states is directly selected as the Gaussian Hidden Markov Model of the target wind turbine, that is, the optimal number of hidden states is selected. Corresponding GHMM: The strong statistical dependence between candidate and target wind turbines indicates that, given the target turbine's power, the conditional power distribution of candidate turbines is more concentrated. Since different operating states typically correspond to different power statistics, it can be further inferred that, given the target turbine's hidden state, candidate turbines tend to fall into a few specific hidden states, thus establishing a relatively stable mapping between their hidden states. Therefore, it can be assumed that the optimal number of hidden states in their corresponding Gaussian Hidden Markov Models (HMMs) is usually quite close. Based on this, the optimal number of hidden states of the target turbine can be directly used as the number of hidden states in the candidate turbine's HMM. Using the same training set partitioning ratio and standardization method as the target turbine, the Baum-Welch algorithm is employed to estimate model parameters on the candidate turbine's training set, thereby constructing the candidate turbine's HMM.

[0055] Specifically, for the candidate wind turbine set any of the wind turbines The power sequence can be preprocessed using the same steps as the target wind turbine, including training-validation set partitioning, standardization, and Gaussian Hidden Markov Model modeling. A model employing hidden state numbers can also be constructed. GHMM: After the above steps, the Gaussian Hidden Markov Model of the target wind turbine and the Gaussian Hidden Markov Model of each candidate wind turbine were obtained.

[0056] Step 206: Calculate the generation mechanism similarity index between each candidate wind turbine and the target wind turbine based on the constructed Gaussian Hidden Markov Model, and select the similar wind turbines in descending order of generation mechanism similarity index.

[0057] In some embodiments, after obtaining the Gaussian Hidden Markov Models for the target wind turbine and each candidate wind turbine, it is necessary to quantify the similarity between each candidate wind turbine and the target wind turbine at the generation mechanism level. Since the statistical scales of power sequences from different wind turbines may differ, directly using the original power sequences for cross-model log-likelihood calculations may result in incomparable likelihood values ​​due to differences in statistical scales between different wind turbines. Therefore, a bidirectional normalization mapping can be performed first: on the one hand, the power sequences of the candidate wind turbines are mapped to the normalization space of the target wind turbine, i.e., the power sequences of the candidate wind turbines are normalized using the mean and standard deviation of the target wind turbine's training set; on the other hand, the power sequences of the target wind turbines are mapped to the normalization space of the candidate wind turbines, i.e., the power sequences of the target wind turbines are normalized using the mean and standard deviation of the candidate wind turbine's training set.

[0058] After completing the bidirectional normalization mapping, the average log-likelihood in both directions is calculated. The first direction involves inputting the candidate wind turbine power sequence mapped to the target wind turbine's normalized space into the target wind turbine's Gaussian Hidden Markov Model (GMM), and calculating the average log-likelihood under that model. This value reflects the explanatory power of the target wind turbine's generative model for the candidate wind turbine's power behavior. The second direction involves inputting the target wind turbine power sequence mapped to the candidate wind turbine's normalized space into the candidate wind turbine's GMM, and calculating the average log-likelihood under that model. This value reflects the explanatory power of the candidate wind turbine's generative model for the target wind turbine's power behavior. The generative mechanism similarity index, due to its directional bias from using only unidirectional log-likelihood, may exhibit a bias. For example, the target wind turbine's GMM may assign a high log-likelihood value to the candidate wind turbine's observation data, while the candidate wind turbine's GMM may not assign a similar log-likelihood value. Therefore, a unidirectional evaluation cannot fully reflect the degree of similarity between the two generative mechanisms. To reduce this directional bias, the average log-likelihood in both directions is symmetricized, and its arithmetic mean is used as an index of the similarity of the generation mechanism between the candidate wind turbine and the target wind turbine.

[0059] For example, for the candidate wind turbine set Any candidate wind turbine , its power sequence Standardized space mapped to the target wind turbine: At the same time, the target wind turbine power sequence Mapped to candidate wind turbines Standardized space: Next, the first mean log-likelihood of the candidate wind turbine in the Gaussian Hidden Markov Model of the target wind turbine is calculated. : Furthermore, the second average log-likelihood of the target wind turbine under the Gaussian Hidden Markov Model of the candidate wind turbines is calculated. : Symmetric mean log-likelihoods are obtained by symmetric means. : In some embodiments, after calculating the generation mechanism similarity index between each candidate wind turbine in the candidate wind turbine set and the target wind turbine, the candidate wind turbines are sorted in descending order of the index. A higher generation mechanism similarity index indicates a greater degree of similarity between the candidate wind turbine and the target wind turbine at the level of power sequence generation mechanism; that is, the corresponding Gaussian Hidden Markov Models of both have stronger explanatory power for each other's observed sequences. Several candidate wind turbines with the highest ranking are selected as the final set of similar wind turbines, which is used for constructing the input features of the subsequent power prediction model.

[0060] For example, according to Sort by size from largest to smallest, and select the top few. Each wind turbine index constitutes a set of wind turbines with a similar generation mechanism. : Therefore, the similar wind turbine screening method provided in this embodiment significantly reduces computational overhead by narrowing the scope of subsequent modeling from all wind turbines to a set of candidate wind turbines through mutual information pre-screening. Furthermore, the optimal number of hidden states for the target wind turbine is determined based on the BIC criterion, and this optimal number of hidden states is uniformly assigned to all candidate wind turbines. While maintaining consistent model complexity, this avoids optimizing the number of hidden states independently for each candidate wind turbine, further improving computational efficiency. Finally, through bidirectional standardized sequence mapping and symmetric average log-likelihood calculation, the mutual interpretability between models is cross-validated from two directions, effectively reducing the directional bias caused by unidirectional evaluation and improving the reliability of similarity determination. The ultimately selected similar wind turbines have a high degree of consistency with the target wind turbine at the power sequence generation mechanism level, providing more valuable input features for subsequent power prediction models.

[0061] In some embodiments, after selecting several similar wind turbines with similar generation mechanisms to the target wind turbine, the historical power data of the target wind turbine and these similar wind turbines can be jointly input into the power prediction model to predict the power value of the target wind turbine at future times. This embodiment provides a residual hedging LSTM network as a power prediction model. This model redesigns the prediction objectives of each layer based on a multi-layer LSTM network, so that the first layer LSTM serves as the baseline prediction branch, outputting the baseline predicted value of the target variable for fitting the true value; the second to Lth layers LSTM serve as the residual prediction branch, used to fit the residual between the true value and the baseline predicted value, thereby improving the accuracy of wind power prediction through the synergistic mechanism of baseline prediction and residual correction.

[0062] For example, the power prediction model includes an L-layer LSTM, where each LSTM layer receives the input sequence sequentially in the time dimension and passes temporal features layer by layer. Specifically, the first LSTM layer directly receives a multivariate time-series input formed by concatenating historical power data of the target wind turbine and similar wind turbines, recursively processes the input sequence in the time dimension, and outputs the final hidden state at the last time step. This final hidden state is then input into the first linear mapping layer, and after linear transformation, a baseline prediction value is output.

[0063] For the second to Lth layers of the LSTM, the input is the hidden state output by the previous LSTM at all time steps, to further extract higher-level temporal representations. Each LSTM outputs its final hidden state at the last time step and inputs it into its respective linear mapping layer, which then outputs the corresponding residual prediction value after a linear transformation. Unlike the first layer of the LSTM, the prediction target of the second to Lth layers of the LSTM is not the actual future power of the target wind turbine, but rather the residual between the actual future power of the target wind turbine and the baseline prediction value.

[0064] Based on this, the final output of the power prediction model is obtained by weighting and adding the baseline prediction value with the residual prediction values ​​of the outputs from the second to the Lth layers.

[0065] In some embodiments, the power prediction model needs to be trained before practical application. The sample data used in the training process comes from the historical power data of the target wind turbine and the power data of similar wind turbines at the corresponding time, wherein the future power of the target wind turbine is used as the prediction target. It should be understood that the prediction object for both the training phase and the prediction phase is the target wind turbine. The difference is that the training phase uses historical data of the target wind turbine and similar wind turbines to construct training samples, while the prediction phase uses historical data of the target wind turbine and similar wind turbines before the time to be predicted as model input to predict the future power of the target wind turbine.

[0066] During training, the target wind turbine's future actual power is used as the supervision signal for the first-layer LSTM, and the baseline prediction loss is calculated. Specifically, the difference between the target wind turbine's future actual power and the baseline prediction value output by the first-layer LSTM is measured using the mean square error method to obtain the baseline prediction loss. Furthermore, the residual between the target wind turbine's future actual power and the baseline prediction value output by the first-layer LSTM is used as the supervision signal for the second to Lth layers of the LSTM, and the corresponding residual prediction loss is calculated.

[0067] During training, the weighted sum of the residual predictions from the second to the Lth LSTM layers is first calculated and added to the baseline prediction to obtain the power prediction result. Then, the parameters of each LSTM layer of the power prediction model are updated based on the baseline prediction loss and the residual prediction losses. Specifically, the parameters of the first LSTM layer are updated based on the baseline prediction loss and the residual prediction losses; the parameters of the second LSTM layer are updated based on the baseline prediction loss and the residual prediction losses. Layer LSTM ( The parameters are based on the first Layer to the first The residual prediction loss corresponding to each layer is updated. Finally, the weights corresponding to the residual prediction branches from the second to the Lth layer are dynamically updated based on the residual prediction loss (note: these weights are used to calculate future power, not the network parameters of the LSTM itself), so that the branches with smaller residual prediction losses receive larger weight values, thereby enhancing the influence of the residual branches with higher prediction accuracy on the final output.

[0068] In some embodiments, the weights corresponding to each residual prediction branch can be updated as follows: First, in a training batch, the prediction error corresponding to each residual prediction branch from the second layer to the Lth layer is calculated. Then, the current weight value corresponding to each residual prediction branch is multiplied by a decay factor determined based on the prediction error of that residual prediction branch to obtain the updated weight value. The decay factor can be constructed using an exponential decay function of the prediction error, such that the decay factor corresponding to a residual prediction branch with a larger prediction error is smaller, resulting in faster weight decay; conversely, the weight decay of a residual prediction branch with a smaller prediction error is slower, resulting in a relatively larger updated weight value. To prevent the weights of some residual prediction branches from decaying to near zero and being completely ignored during multiple updates, a preset lower bound constraint can be applied to the updated weight value to ensure that each residual prediction branch retains a certain weight.

[0069] Therefore, the weights corresponding to each residual prediction branch from the second to the Lth layer are continuously and adaptively adjusted during the training process, so that the residual prediction branch with smaller prediction error receives a larger weight, while the weights corresponding to the residual prediction branch with larger prediction error gradually decrease but do not decay to zero. In this way, while retaining the effective information of each residual prediction branch, the adaptive combination of prediction results of each layer is achieved.

[0070] In some embodiments, during backpropagation, the parameters of the first LSTM layer are updated based on the baseline prediction loss and the prediction losses of each residual; Layer LSTM ( The parameters are based on the first Layer to the first The residual prediction loss corresponding to each layer is updated. Unlike existing hedged LSTMs, in this embodiment, the supervision signal for each residual prediction branch from the second to the Lth layer is the residual between the actual future power of the target wind turbine and the baseline prediction value output by the first-layer LSTM. Therefore, the second to the Lth layers of the LSTM no longer repeatedly predict the target variable, but instead uniformly learn the mapping relationship between the input time-series features and the prediction residuals, and output the corresponding residual prediction values ​​to correct the baseline prediction value output by the first-layer LSTM.

[0071] Therefore, the power prediction model provided in this embodiment redesigns the prediction objectives of each layer based on a multi-layer LSTM network. The first-layer LSTM outputs the baseline prediction value of the target variable, which is used to fit the true value; the second to Lth layers of LSTM are used to fit the residual between the true value and the baseline prediction value. The residual prediction loss of the second to Lth layers participates in the parameter optimization of the first-layer LSTM through gradient backpropagation, so that the gradient information from the residual correction task can guide the parameter update direction of the first layer. This allows the temporal features extracted by the first-layer LSTM to not only be suitable for baseline prediction, but also better support the subsequent residual correction task, thereby achieving collaborative optimization between the baseline prediction layer and the residual correction layer, which is one of the important reasons for the improvement in model prediction performance.

[0072] In some embodiments, the input features of multiple wind turbines can be constructed first. The target wind turbine and the turbines selected through the aforementioned steps are then obtained. Historical power data of similar wind turbines. To ensure the validity of the model input and meet the rated power limit of the wind turbines, boundary processing can be performed on the dataset: null values ​​in the data are uniformly set to 0, and power values ​​exceeding 1500kW are truncated to 1500kW. Subsequently, the dataset is divided into training, validation, and test sets in chronological order. Let the lengths of the three datasets be respectively... , and Any dataset can be presented in the form of a power matrix: in, , Represents wind turbine At time step The power value.

[0073] To meet the structural requirements of the model's input-output, a sample-label tensor pair is constructed based on a sliding window mechanism. Let the length of the historical window be... Construct sample-label tensors : in, , No. Sample for: in, ,That .

[0074] Based on this, the sample tensor of the training set Constructing the Z-score normalization operator Min-Max normalization operator is constructed based on the label tensor of the training set. Furthermore, the above operators are applied uniformly to the training set, validation set, and test set to ensure the consistency of data distribution: in, , This represents the corresponding standardized sample-label tensor pair.

[0075] Figure 3 This is a schematic diagram of the structure of a power prediction model provided in an exemplary embodiment.

[0076] Please refer to Figure 3 The power prediction model comprises L layers of LSTM, each layer receiving the input sequence sequentially in the time dimension and passing temporal features layer by layer. At each time step, the... Layer LSTM receives the first The hidden states output by the LSTM layer are used to extract temporal features layer by layer. Representing the LSTM in the first layer The hidden state at each time step, i.e., the feature representation output by the LSTM network at the current moment, is adopted. Representing the Layer LSTM in the first The cell state at each time step, i.e. the long-term memory state maintained within that LSTM layer, can then be represented as: in, Representing the A layered LSTM is a full sequence state mapping formed by the recursion of state transition functions in the time dimension. Represents the original input sequence. Representing the The initial state of the LSTM layer.

[0077] The first LSTM layer serves as the baseline prediction layer of the network, by analyzing its final hidden state. Perform a linear mapping to obtain a baseline prediction of the target wind turbine's future power. : in, This represents the linear mapping corresponding to the first LSTM layer. This LSTM layer belongs to the baseline prediction branch, which directly learns the main variation law of the target wind turbine power and undertakes the main prediction task.

[0078] The second to Lth layers of the LSTM form the residual prediction branch. The final hidden states of each LSTM layer are input into the linear mapping layer to obtain the corresponding residual prediction values. in, Representing the The output of the LSTM layer, Representative on the first The final hidden state of the LSTM layer A linear mapping is performed using a linear transformation. Therefore, from the second to the Lth layer of the LSTM, the target power is no longer directly predicted; instead, the true power is learned. Residuals between the baseline forecast and the baseline forecast : Therefore, the shallow network of the power prediction model is responsible for fitting the main trend information in the power sequence, while the deep network is responsible for learning the remaining error components after the shallow prediction, realizing the functional division of different levels, thereby avoiding repeated modeling of the same prediction target by each layer.

[0079] The outputs of the LSTM layers 2 through L are weighted and combined using hedging weights, resulting in the final output. for: in, Representing the The weights corresponding to the Lth LSTM layers are determined by this formula, and the reliability of the Lth LSTM layers is determined by their respective weight coefficients. Adjustments are made to achieve adaptive integration of prediction results from each layer.

[0080] Figure 3 The HBP (Hedge Backpropagation) module in the code represents the joint update process of network parameters and hedging weights, as shown in the following formula.

[0081] In the One batch, first layer and The backpropagation process is as follows: For the second to the Lth layer LSTM network, the predicted target lies in the residual space. and The parameter updates are affected by the combined residual prediction errors of this layer and subsequent layers. The first batch Layer LSTM network and The backpropagation process is as follows: In the two formulas above, Representing the linear layer in each batch The parameters, Representing the The first batch Parameters of a multi-layer LSTM network, , This represents the total number of batches included in each training epoch. Represents the learning rate. Represents the loss function. Representing the The weights corresponding to the LSTM layer.

[0082] During training, weights can be dynamically updated based on the prediction error of each layer, making the reliability of the better-performing residual branches higher. Weights The update rules are as follows: in, This is the weight decay rate. This smoothing coefficient is applied to the updated weights to prevent the weights of certain layers from decaying to near zero and being completely discarded after multiple updates, serving as a constraint. The lower bound.

[0083] Therefore, the weights corresponding to each residual prediction branch from the second to the Lth layer are continuously and adaptively adjusted during the training process, so that the residual prediction branch with smaller prediction error receives a larger weight, while the weights corresponding to the residual prediction branch with larger prediction error gradually decrease but do not decay to zero. In this way, while retaining the effective information of each residual prediction branch, the adaptive combination of prediction results of each layer is achieved.

[0084] After model training is complete, prediction can be performed. The historical power data of the target wind turbine and similar wind turbines are preprocessed using the same standardization method as in the training phase, and the resulting feature tensors are then analyzed. Input the trained power prediction model above The predicted power of the target wind turbine at future times is obtained: in, This represents the predicted future power of the target wind turbine at a future time.

[0085] Since the prediction results output by the power prediction model are located in the normalized space, the maximum and minimum values ​​of the labels recorded during the training phase can be used to perform an inverse normalization transformation on the prediction results, thereby restoring them to the actual power scale and obtaining the power prediction value of the target wind turbine at future times (in kilowatts).

[0086] In some embodiments, the mean error and extreme error are characterized by MAE (Mean Absolute Error) and RMSE (Root Mean Square Error), respectively. The above normalization process can be expressed as: in, and Let be the tensors composed of predicted power and actual power, respectively. For a single-step prediction, their _i_ and _v_i_ are tensors. The samples are respectively and The model's MAE and RMSE are as follows: In a typical application scenario, this method can be deployed on a wind farm's station control server or edge computing node. The data acquisition and monitoring system within the wind farm collects power data from each wind turbine in real time and transmits the data to the station control server via industrial Ethernet or fieldbus. The station control server runs a pre-trained power prediction model. Whenever new data arrives, the server extracts historical window data from the target turbine and similar turbines, executes the aforementioned prediction process, and can output the predicted power value of the target turbine at a future time within seconds. This predicted value can be sent to the wind farm's energy management system, providing operational decision-making references for dispatchers and providing data support for optimized active power scheduling and automatic generation control strategy formulation within the wind farm.

[0087] In another application scenario, this method can be deployed on a cloud-based prediction platform. The wind farm uploads historical and real-time operational data collected from each wind turbine to the cloud via a data upload gateway. The cloud server then uses a trained power prediction model to perform batch or on-demand predictions and returns the prediction results to the wind farm's local system via an application programming interface (API). Cloud deployment facilitates centralized management of prediction tasks across multiple wind farms and allows for dynamic allocation of computing resources based on the scale of the prediction task, improving the processing efficiency of large-scale prediction tasks. Those skilled in the art can flexibly choose a suitable deployment method based on the wind farm's network conditions, computing resource distribution, and real-time requirements.

[0088] To verify the effectiveness of the power prediction model described in this specification, this embodiment uses actual operating data from a wind farm to conduct an experiment.

[0089] This embodiment selects an ultra-short-term power prediction mode for experimental analysis: using the power data of the target wind turbine and similar wind turbines obtained through screening over the past 2 hours (i.e., 12 time steps, each step 10 minutes) as input, the power of the target wind turbine in the next 10 minutes (i.e., 1 time step) is predicted.

[0090] To avoid bias from a single evaluation metric, this embodiment uses mean absolute error and root mean square error as prediction performance evaluation metrics, respectively, to comprehensively evaluate the overall prediction error level of the model and its sensitivity to larger prediction errors. During the testing phase, the input power tensor is first standardized, then the prediction results in the standardized space are obtained through model forward propagation, and finally restored to the original physical scale through inverse normalization transformation.

[0091] To verify the superiority of the power prediction model provided in this specification, the following comparative model is set up in this embodiment: (1) One to five layers of LSTM (referred to as 1-LSTM, 2-LSTM, 3-LSTM, 4-LSTM, and 5-LSTM respectively), and the SGD algorithm is used for parameter optimization. (2) A hedging LSTM network (denoted as H-LSTM-NL) with a nonlinear mapping mechanism at the output end. This model is built on a multi-layer LSTM network and introduces a cascaded structure consisting of a nonlinear feature transformation layer and a linear mapping layer at the output end. Under the experimental conditions of this embodiment, 5-LSTM exhibits the most obvious performance degradation. Therefore, this embodiment uses 5-LSTM as a basis and introduces a hedging algorithm to construct H-LSTM-NL; (3) The output end adopts a hedging LSTM network (denoted as H-LSTM-L). This model removes the nonlinear feature transformation layer at the output end from the H-LSTM-NL model and retains only the linear mapping layer.

[0092] During the training phase, to ensure comparability between different models, some structures and training parameters were uniformly set (see Table 1), and corresponding unique parameters were set for different models (see Table 2).

[0093] Table 1 Shared parameter settings for each model Table 2. Specific parameter settings for each model Table 3 Comparison of wind turbine power prediction performance between the power prediction model provided in this application and the comparative model As shown in Table 3: First, the MAE and RMSE of LSTM layers 1 to 5 gradually increase with the number of layers. This result indicates that, under the current experimental conditions, as the number of network layers increases, the model may experience performance degradation due to problems such as difficulty in gradient propagation, increased risk of overfitting, and redundant feature representation, i.e., a depth degradation effect.

[0094] Second, the prediction performance of H-LSTM-NL (MAE: 72.87, RMSE: 94.49) is significantly better than that of 5-LSTM (MAE: 145.97, RMSE: 176.79), indicating that the introduction of the hedging algorithm can effectively alleviate the deep degradation effect.

[0095] Third, the performance of H-LSTM-L (MAE: 60.01, RMSE: 82.27) further outperforms H-LSTM-NL (MAE: 72.87, RMSE: 94.49), and surpasses the best-performing 1-LSTM (MAE: 69.63, RMSE: 89.30). This indicates that, under current experimental conditions, the linear mapping mechanism is more conducive to obtaining better prediction performance.

[0096] Fourth, the power prediction model provided in this specification achieved the best prediction performance (MAE: 41.93, RMSE: 71.33). This result shows that, under the current experimental conditions, the prediction mechanism based on the joint optimization of baseline prediction and residual correction can further improve the prediction accuracy of the model.

[0097] Figure 4 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 4 At the hardware level, the device includes a processor 402, an internal bus 404, a network interface 406, memory 408, and non-volatile memory 410, and may also include other hardware required for implementation. One or more embodiments of this specification can be implemented in software. For example, the processor 402 can read a stored computer program from the non-volatile memory 410 and load it into memory 408 for execution. Of course, in addition to software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic device implementation or a combination of software and hardware implementation. That is to say, the execution entity of the above processing flow is not limited to software functional modules, but can also be implemented by hardware modules or software and hardware combined modules.

[0098] Please refer to Figure 5 The wind power prediction device 500 can be applied to, for example... Figure 4 The device shown implements the technical solution described in this specification. The wind power prediction device 500 may include: Data acquisition unit 501 acquires the historical power sequence of the target wind turbine and other wind turbines in the same wind farm; The similar wind turbine screening unit 502 performs probabilistic modeling of the power sequence of the target wind turbine and each candidate wind turbine selected from the other wind turbines based on the Gaussian Hidden Markov Model, and evaluates the similarity of the generation mechanism between the target wind turbine and each candidate wind turbine based on the similarity between the models, so as to screen out a number of similar wind turbines with similar generation mechanisms to the target wind turbine from the candidate wind turbines. The power prediction unit 503 inputs the historical power data of the target wind turbine and the similar wind turbines into the power prediction model to obtain the future power prediction value of the target wind turbine. The power prediction model includes an L-layer LSTM, wherein the first layer LSTM serves as the baseline prediction branch, which is used to fit the future actual power of the target wind turbine and output the baseline prediction value; the second to Lth layers LSTM serve as the residual prediction branch, which are used to fit the residual between the future actual power of the target wind turbine and the baseline prediction value and output the residual prediction value; the future power prediction value is calculated based on the baseline prediction value and the residual prediction value.

[0099] Optionally, the step of probabilistically modeling the power sequences of the target wind turbine and each candidate wind turbine selected from the other wind turbines based on a Gaussian Hidden Markov Model, and evaluating the similarity of the generation mechanisms of the target wind turbine and each candidate wind turbine based on the similarity between the models, so as to select a number of similar wind turbines from the candidate wind turbines that have similar generation mechanisms to the target wind turbine, includes: Calculate the mutual information between the historical power sequence of the target wind turbine and the historical power sequences of other wind turbines in the same wind farm, and select several other wind turbines as candidate wind turbines based on the mutual information estimate. The optimal number of hidden states in the Gaussian Hidden Markov Model of the target wind turbine is determined, and the optimal number of hidden states is used as the number of hidden states in the Gaussian Hidden Markov Model of each candidate wind turbine, so as to construct the Gaussian Hidden Markov Model of the target wind turbine and each candidate wind turbine respectively. The generation mechanism similarity index between each candidate wind turbine and the target wind turbine is calculated based on the constructed Gaussian Hidden Markov Model, and the similar wind turbines are selected in descending order of generation mechanism similarity index.

[0100] Optionally, constructing Gaussian Hidden Markov Models for the target wind turbine and each candidate wind turbine includes: For the target wind turbine, the Gaussian Hidden Markov Model with estimated parameters corresponding to the optimal number of hidden states is selected as the Gaussian Hidden Markov Model of the target wind turbine. For any of the candidate wind turbines, the model parameters are estimated using the Baum-Welch algorithm based on the optimal number of hidden states to construct a Gaussian hidden Markov model for the candidate wind turbine.

[0101] Optionally, the step of calculating the generation mechanism similarity index between each candidate wind turbine and the target wind turbine based on the constructed Gaussian Hidden Markov Model includes: The power sequence of the candidate wind turbines is mapped to the normalized space of the target wind turbine, and the first average log-likelihood is calculated under the Gaussian hidden Markov model of the target wind turbine. The power sequence of the target wind turbine is mapped to the normalized space of the candidate wind turbine, and the second average log-likelihood is calculated under the Gaussian hidden Markov model of the candidate wind turbine. The mean of the first average log-likelihood and the second average log-likelihood is calculated and used as an index of the generation mechanism similarity between the candidate wind turbine and the target wind turbine.

[0102] Optionally, the future power prediction is obtained by weighting and adding the baseline prediction with the residual predictions of the LSTM outputs from the second to the Lth layers.

[0103] Optionally, the training process of the power prediction model includes: The baseline prediction loss is calculated using the actual power value of the target wind turbine as the supervision signal of the first layer LSTM. The difference between the actual power value of the target wind turbine and the baseline prediction value output by the first layer LSTM is used as the supervision signal of the second to Lth layers LSTM, and the residual prediction loss of the second to Lth layers LSTM is calculated. The weights corresponding to the residual prediction branches from the second to the Lth layer are dynamically updated based on the residual prediction loss, so that the branches with smaller residual prediction losses receive larger weight values. Calculate the weighted sum of the residual prediction values ​​and corresponding weight values ​​of the LSTM output from the second to the Lth layer, and add it to the baseline prediction value to obtain the power prediction result; The parameters of each LSTM layer of the power prediction model are updated based on the baseline prediction loss and each residual prediction loss, wherein the parameters of the first LSTM layer are updated based on the baseline prediction loss and each residual prediction loss; The parameters of the LSTM layer are based on the first layer. Layer to the first The residual prediction loss corresponding to the layer is updated, where, The value of is greater than or equal to 2 and less than or equal to L.

[0104] Optionally, dynamically updating the weights corresponding to the residual prediction branches from the second to the Lth layer based on the residual prediction loss includes: Calculate the prediction error for each residual prediction branch from the second layer to the Lth layer respectively; The weights corresponding to each residual branch are multiplied by a decay factor determined based on the prediction error of each residual branch to obtain updated weight values, wherein the decay factor is an exponential decay function of the prediction error.

[0105] Based on the same concept as the methods described above, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described in any of the above embodiments by executing the executable instructions.

[0106] Based on the same concept as the methods described above, this specification also provides a computer-readable storage medium having computer instructions stored thereon that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0107] Based on the same concept as the methods described above, this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

Claims

1. A method for predicting wind power output, characterized in that, The method includes: Obtain the historical power sequence of the target wind turbine and other wind turbines in the same wind farm; Based on the Gaussian Hidden Markov Model, the power sequence of the target wind turbine and each candidate wind turbine selected from the other wind turbines is probabilistically modeled, and the similarity of the generation mechanism between the target wind turbine and each candidate wind turbine is evaluated according to the similarity between the models, so as to select a number of similar wind turbines with similar generation mechanism to the target wind turbine from the candidate wind turbines. The historical power data of the target wind turbine and the similar wind turbines are input into the power prediction model to obtain the future power prediction value of the target wind turbine; The power prediction model includes an L-layer Long Short-Term Memory (LSTM) system. The first LSTM layer serves as the baseline prediction branch, which is used to fit the future actual power of the target wind turbine and output the baseline prediction value. The second to Lth LSTM layers serve as the residual prediction branches, which are used to fit the residual between the future actual power of the target wind turbine and the baseline prediction value and output the residual prediction value. The future power prediction value is calculated based on the baseline prediction value and the residual prediction value.

2. The method according to claim 1, characterized in that, The method involves probabilistically modeling the power sequences of the target wind turbine and each candidate wind turbine selected from the other wind turbines based on a Gaussian Hidden Markov Model, and evaluating the similarity of the generation mechanisms of the target wind turbine and each candidate wind turbine based on the similarity between the models, in order to select several similar wind turbines from the candidate wind turbines that have similar generation mechanisms to the target wind turbine, including: Calculate the mutual information between the historical power sequence of the target wind turbine and the historical power sequences of other wind turbines in the same wind farm, and select several other wind turbines as candidate wind turbines based on the mutual information estimate. The optimal number of hidden states in the Gaussian Hidden Markov Model of the target wind turbine is determined, and the optimal number of hidden states is used as the number of hidden states in the Gaussian Hidden Markov Model of each candidate wind turbine, so as to construct the Gaussian Hidden Markov Model of the target wind turbine and each candidate wind turbine respectively. The generation mechanism similarity index between each candidate wind turbine and the target wind turbine is calculated based on the constructed Gaussian Hidden Markov Model, and the similar wind turbines are selected in descending order of generation mechanism similarity index.

3. The method according to claim 2, characterized in that, The construction of Gaussian Hidden Markov Models for the target wind turbine and each candidate wind turbine includes: For the target wind turbine, the Gaussian Hidden Markov Model with estimated parameters corresponding to the optimal number of hidden states is selected as the Gaussian Hidden Markov Model of the target wind turbine. For any of the candidate wind turbines, the model parameters are estimated using the Baum-Welch algorithm based on the optimal number of hidden states to construct a Gaussian hidden Markov model for the candidate wind turbine.

4. The method according to claim 2, characterized in that, The calculation of the generation mechanism similarity index between each candidate wind turbine and the target wind turbine based on the constructed Gaussian Hidden Markov Model includes: The power sequence of the candidate wind turbines is mapped to the normalized space of the target wind turbine, and the first average log-likelihood is calculated under the Gaussian hidden Markov model of the target wind turbine. The power sequence of the target wind turbine is mapped to the normalized space of the candidate wind turbine, and the second average log-likelihood is calculated under the Gaussian hidden Markov model of the candidate wind turbine. The mean of the first average log-likelihood and the second average log-likelihood is calculated and used as an index of the generation mechanism similarity between the candidate wind turbine and the target wind turbine.

5. The method according to claim 1, characterized in that, The future power prediction is obtained by weighting and adding the baseline prediction to the residual predictions output by the LSTMs from the second to the Lth layers.

6. The method according to claim 1, characterized in that, The training process of the power prediction model includes: The baseline prediction loss is calculated using the actual power value of the target wind turbine as the supervision signal of the first layer LSTM. The difference between the actual power value of the target wind turbine and the baseline prediction value output by the first layer LSTM is used as the supervision signal of the second to Lth layers LSTM, and the residual prediction loss of the second to Lth layers LSTM is calculated. The weights corresponding to the residual prediction branches from the second to the Lth layer are dynamically updated based on the residual prediction loss, so that the branches with smaller residual prediction losses receive larger weight values. Calculate the weighted sum of the residual prediction values ​​and corresponding weight values ​​of the LSTM output from the second to the Lth layer, and add it to the baseline prediction value to obtain the power prediction result; The parameters of each LSTM layer of the power prediction model are updated based on the baseline prediction loss and each residual prediction loss, wherein the parameters of the first LSTM layer are updated based on the baseline prediction loss and each residual prediction loss; The parameters of the LSTM layer are based on the first layer. The residual prediction loss from layer L to layer L is updated, where, The value of is greater than or equal to 2 and less than or equal to L.

7. The method according to claim 6, characterized in that, The step of dynamically updating the weights corresponding to the residual prediction branches from the second layer to the Lth layer based on the residual prediction loss includes: Calculate the prediction error of the residual prediction value for each residual prediction branch from the second layer to the Lth layer respectively; The weights corresponding to each residual branch are multiplied by a decay factor determined based on the prediction error of each residual branch to obtain updated weight values, wherein the decay factor is an exponential decay function of the prediction error.

8. A wind power prediction device, characterized in that, The device includes: The data acquisition unit acquires the historical power sequence of the target wind turbine and other wind turbines in the same wind farm. The similar wind turbine screening unit performs probabilistic modeling of the power sequence of the target wind turbine and each candidate wind turbine selected from the other wind turbines based on the Gaussian Hidden Markov Model, and evaluates the similarity of the generation mechanism between the target wind turbine and each candidate wind turbine based on the similarity between the models, so as to screen out a number of similar wind turbines with similar generation mechanisms to the target wind turbine from the candidate wind turbines. The power prediction unit inputs the historical power data of the target wind turbine and the similar wind turbines into the power prediction model to obtain the future power prediction value of the target wind turbine. The power prediction model includes an L-layer LSTM, wherein the first layer LSTM serves as the baseline prediction branch, used to fit the future actual power of the target wind turbine and output the baseline prediction value; the second to Lth layers of LSTM are used to fit the residual between the future actual power of the target wind turbine and the baseline prediction value, and output the residual prediction value; the future power prediction value is calculated based on the baseline prediction value and the residual prediction value.

9. An electronic device, comprising: processor; A memory for storing processor-executable instructions; wherein the processor implements the steps of the method as claimed in any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having stored thereon computer instructions that, when executed by a processor, implement the steps of the method as claimed in any one of claims 1 to 7.

Images