Photovoltaic power generation prediction method and system based on multi-data model fusion
Through reinforcement learning and multi-data model fusion method, using Fisher information and other algorithms to adjust the model weights in real time, the problem of insufficient accuracy and reliability in photovoltaic power prediction is solved, and higher prediction accuracy and reliability are achieved sex.
Patent Information
- Application Number
- CN202111004446.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-30
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-08-30
AI Technical Summary
The existing photovoltaic power prediction methods have insufficient accuracy and reliability in ultra-short-term predictions, especially the traditional multi-data model fusion method cannot dynamically select the optimal weight, resulting in the fusion value not reaching the optimal value.
Using reinforcement learning method, using Fisher information as a reward, the weights of each model are adjusted in real time for online fusion, and combining local outlier factor algorithm, multi-layer perceptron regression algorithm, feature selection and Q-learning algorithm to realize multi-data model fusion prediction.
This significantly reduces the error index, improves prediction accuracy and reliability, and makes the prediction of photovoltaic power generation more accurate and reliable.
Smart Images

Figure CN113919545B_ABST
Abstract
Description
Technical Field
[0001] The present invention is applied to photovoltaic power generation prediction, and relates to a multi-model fusion ultra-short-term photovoltaic power generation prediction method and system. Background Art
[0002] As a new type of clean energy generation technology, photovoltaic power generation has played a positive role in promoting the diversification of energy supply and alleviating the shortage of power supply points. With the rapid development and large-scale application of photovoltaic power generation technology, the problems caused by its grid connection need to be solved urgently. Because photovoltaic power generation is easily affected by the geographical environment, solar radiation and various meteorological factors, it has significant volatility and intermittency. If large-scale grid connection is directly carried out, the network structure, flow distribution and operation mode of the power grid will change, causing power fluctuations in the power grid, affecting the stability of the power grid and scheduling planning. Therefore, the power generation power prediction of photovoltaic power generation is particularly important. Accurate prediction of photovoltaic power generation is conducive to the grid dispatchers to reasonably adjust the power generation plan and maintain the stable operation of the power system.
[0003] At present, in order to improve the accuracy and reliability of model prediction, researchers have proposed the method of multi-data model fusion. Compared with the single model method, the multi-data model fusion method can effectively make up for the shortcomings of each single model method and make the prediction results more reliable. In order to reduce the large errors in the predicted values, the traditional multi-data model fusion method mainly uses the weights of each sub-model based on the equal weight average method, the mean absolute error method, the MAPE-RW method, the entropy weight method and the inverse error square sum method to calculate the weights of each sub-model and perform weighted fusion methods. Although this type of method improves the weight calculation method according to the overall prediction effect of each model, so that the accuracy of the fusion prediction result is improved, it does not consider the fusion effect of the calculated weights on the local prediction values of each model. In addition, when the above methods fuse the models, their weights are fixed values, and the optimal weights cannot be dynamically selected during prediction. The model with the smallest overall error may not have the smallest local error. According to the overall prediction effect or the average distribution of weights, the final fusion value will not reach the optimal value. In this case, reinforcement learning can be used to solve these problems. As one of the most active research areas in artificial intelligence, reinforcement learning has great advantages in solving learning and optimization problems. It obtains feedback information through exploration and interaction with the environment, and adjusts strategies for learning and optimization based on the feedback information. After continuous self-optimization, it gradually approaches the optimal solution. Therefore, the reinforcement learning method is used to find the optimal weights of each sub-model for fusion, so as to achieve the purpose of improving the accuracy of photovoltaic power generation prediction. Summary of the invention
[0004] In order to overcome the shortcomings of the existing ultra-short-term photovoltaic power prediction, which is not accurate and reliable enough, the present invention proposes a photovoltaic power prediction method based on the fusion of multiple data models. The method adopts the reinforcement learning method, takes Fisher information as a reward, determines the optimal weight according to the real-time prediction results of each model, and performs online fusion of the prediction results of each model. The method can adjust the weight of each model in real time to improve the accuracy of the prediction results.
[0005] The technical solution adopted by the present invention is:
[0006] A photovoltaic power generation prediction method based on multi-data model fusion is proposed. The multi-data model fusion is performed by reinforcement learning method to realize ultra-short-term photovoltaic power generation prediction and further improve its accuracy. The method specifically includes the following steps:
[0007] Step 1: Data preprocessing:
[0008] Step 1-1: Use the local outlier factor (LOF) algorithm to detect and remove outliers. The degree of outlier is evaluated by calculating the outlier factor of each power point. First, the Euclidean distance between each power point and other power points is calculated. Let k∈N+, and the power point is x i , power point x i and the kth point closest to it The distance between them is denoted as d k (x i ), then the power point x i The reachable distance to all power points within its k-distance neighborhood is:
[0009] d r (x i ,x i o )=max{d k (x i o ),d(x i ,x i o )} (1)
[0010] In the formula, Represents x i A power point within the k-distance neighborhood; Represents x i and The Euclidean distance between .
[0011] Then calculate the local reachable density of each power point, and the mathematical formula is expressed as:
[0012]
[0013] Finally, the outlier factor of the power point is calculated, and the mathematical formula is expressed as:
[0014]
[0015] If the local outlier factor of a power point is close to 1, it means that the power point is slightly different from its neighborhood and is considered a normal point; if the local outlier factor of a power point is much greater than 1, it means that the power point is significantly different from its neighborhood and is considered an outlier. This algorithm is used to detect and eliminate outliers in photovoltaic power generation.
[0016] Step 1-2: Use the multi-layer perceptron regression algorithm to fill in the missing data. The non-missing and missing data of photovoltaic power generation are used as training sets and test sets respectively. The multi-layer perceptron regression algorithm is used to train in the training set, and then the trained model is used to predict the test set, and finally the prediction results are used to fill in the missing data. The multi-layer perceptron model consists of an input layer, a hidden layer, and an output layer. The mathematical formula for each node from the input layer to the hidden layer and then to the output layer is expressed as:
[0017] x ij =g(W i X i-1 +b i-1 ) (4)
[0018] Where x ij The value of the jth neuron in the i-th layer, W i represents the weight vector of the jth neuron in the i-1th layer to the i-th layer; X i-1 b represents the vector of all neuron values in the i-1th layer; i-1 represents the bias term of the i-1th layer; g represents the activation function, and the ReLU function is used as the activation function.
[0019] The loss function of the multilayer perceptron regression algorithm is expressed in mathematical formula as follows:
[0020]
[0021] In the formula, Represents the output value of the multilayer perceptron regression algorithm; y i Represents the true value, and the multilayer perceptron regression algorithm stops calculating when the loss function value converges.
[0022] Step 2: Feature selection. By selecting representative features from multiple features through feature selection and reducing irrelevant feature information, the difficulty of model learning can be reduced and the training efficiency can be improved. The present invention adopts the distance correlation coefficient method to determine the correlation between each feature and the power generation power, and selects the features with a large correlation with the power generation power. Taking instantaneous radiation, scattered radiation, normal radiation, air temperature, air pressure, humidity and backplane temperature as features, and photovoltaic power generation power as a label, the Pearson correlation coefficient and distance correlation coefficient of each feature and the label are calculated, and the mathematical formula is expressed as follows:
[0023]
[0024] In the formula, x represents a certain feature, y represents the power generation, and v 2 (x,x) represents the feature distance square; v 2 (y,y) represents the distance variance of the generated power; v 2 (x,y) represents the distance covariance between the feature and the generated power.
[0025] The distance correlation coefficient D(x,y)∈[0,1] between each feature and photovoltaic power generation is obtained by calculation. The closer it is to 1, the stronger the correlation between the selected feature, label and power generation; the closer it is to 0, the weaker the correlation between the selected feature and power generation.
[0026] Step 3: Train sub-models. Divide the data into training set, validation set and test set. Train the support vector machine regression (SVR), multivariate linear regression (MLR) and Bayesian ridge regression (BRR) models in the training set, and verify the trained models in the validation set. Select the best models as sub-models.
[0027] Support vector machine regression (SVR) is a machine learning method based on the risk minimization criterion to solve regression fitting problems. The principle of the algorithm is to perform linear regression in high-dimensional space to find an optimal hyperplane so that the total deviation of the sample points from the hyperplane is minimized. Its mathematical formula is expressed as:
[0028]
[0029] In the formula, ω represents the weight vector, represents the kernel function, and b represents the bias.
[0030] Multiple linear regression (MLR) is analyzed by establishing a mathematical model between the dependent variable and its multiple independent variables. The mathematical formula of the multiple linear regression model with independent variables is expressed as:
[0031] y=β0+β1x1+...+β m x m +ε (8)
[0032] Where β0,β1,...β m represents the regression coefficient; y represents the dependent variable; x0, x1, ...x m represents the independent variable; ε represents the random error.
[0033] Bayesian Ridge Regression (BRR) is a regression algorithm based on Bayesian theory. It introduces the L2 regularization term in the estimation process, so it is called Bayesian Ridge Regression. The mathematical formula of Bayesian linear regression is:
[0034]
[0035] In the formula, m represents the dimension of the sample space; ω represents the parameter vector; x represents the input vector; represents a nonlinear function of x.
[0036] Step 4: Multi-data model fusion prediction. Use the best sub-models obtained through training to predict the data in the test set, and use the Q-learning algorithm to fuse the prediction results of each sub-model. The fusion process can be expressed as a Markov decision process {S, A, R}, where S represents the state set, A represents the action set, and R represents the reward function, where the state and action are a set of weights ω1, ω2, ..., ω n , record the state ω at time t s1 ,ω s2 ,...,ω sn for Remember the action at time t a1 ,ω a2 ,...,ω an for The reward function is the Fisher information, which can reflect the accuracy of parameter estimation. The larger the Fisher information, the higher the accuracy of parameter estimation. The calculation process of Fisher information is as follows.
[0037] First, we obtain the mathematical expression of the model prediction value:
[0038] M i =M0+ΔM i (10)
[0039] Where M i represents the prediction power of the ith model, M0 represents the true value, ΔM i represents the uncertainty of the power predicted by the ith model. Assume in represents the variance of the prediction power of the ith model.
[0040] The predicted power of the model conforms to the Gaussian distribution, that is:
[0041]
[0042] Since all models are independent, their joint probability density is:
[0043]
[0044] Where, Γ=(M1,M2,...M n ).
[0045] Finally, the Fisher information is calculated, and its mathematical expression is:
[0046]
[0047] In the fusion process, a state is obtained from the state set and an action is selected from the action set. After the action is taken, the fusion system gives a reward value according to the reward function and switches to the next state. The fusion system learns repeatedly until it reaches the terminal state and the Q value converges to the maximum value. Its update mechanism is expressed by the Bellman equation:
[0048]
[0049] In the formula, represents the Q value before the update at time t; represents the updated Q value at time t; Indicates the current status; Indicates the action taken in the current state; Indicates the next state; Indicates status The actions that can be taken under the condition of , δ represents the learning rate, which controls the ratio of new and old experiences in training; γ represents the discount factor, which indicates the importance of future rewards.
[0050] A system for implementing a photovoltaic power generation prediction method of a multi-data model fusion of the present invention comprises a data preprocessing module, a feature selection module, a sub-model training module, and a multi-data model fusion prediction module connected in sequence, wherein:
[0051] The data preprocessing module uses the local outlier factor (LOF) algorithm to detect and remove outliers, and the multi-layer perceptron regression algorithm to fill in missing data.
[0052] The feature selection module selects representative features from multiple features through feature selection, reduces irrelevant feature information, reduces the difficulty of model learning, and improves training efficiency.
[0053] The sub-model training module divides the data into training set, validation set and test set. Support vector machine regression (SVR), multivariate linear regression (MLR) and Bayesian ridge regression (BRR) models are trained in the training set, and the trained models are verified in the validation set. The best models are selected as sub-models.
[0054] The multi-data model fusion prediction model uses the best sub-models obtained through training to predict the data in the test set, and uses the Q-learning algorithm to fuse the prediction results of each sub-model.
[0055] Compared with the prior art, the present invention has the following beneficial effects:
[0056] 1. Compared with the single model prediction method, the prediction method of multi-data model fusion proposed in the present invention has significantly reduced error indicators and higher prediction accuracy;
[0057] 2. Compared with other multi-data model fusion prediction methods, the multi-data model fusion prediction method proposed in the present invention can determine the optimal weight according to the real-time prediction results of each model, and perform online fusion of the prediction results of each model, with lower error indicators, higher prediction accuracy and stronger reliability;
[0058] 3. The photovoltaic power generation prediction method based on multi-data model fusion proposed in the present invention has certain practicality. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 It is a characteristic correlation analysis diagram of the present invention;
[0060] Figure 2 is a flow chart of the Q-learning algorithm of the present invention;
[0061] Figure 3 It is a schematic diagram of the multi-data model fusion process of the present invention;
[0062] Figure 4 This is a comparison chart of the indicator evaluation of the present invention, a single model and other fusion methods. DETAILED DESCRIPTION
[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention is clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0064] Reference Figure 1 to Figure 4A photovoltaic power generation prediction method based on multi-data model fusion, the specific steps are as follows:
[0065] Step 1: Data preprocessing:
[0066] Step 1-1: Use the local outlier factor (LOF) algorithm to detect and remove outliers. The degree of outlier is evaluated by calculating the outlier factor of each power point. First, the Euclidean distance between each power point and other power points is calculated. Let k∈N+, and the power point is x i , power point x i and the kth point closest to it The distance between them is denoted as d k (x i ), then the power point x i The reachable distance to all power points within its k-distance neighborhood is:
[0067] d r (x i ,x i o )=max{d k (x i o ),d(x i ,x i o )} (1)
[0068] In the formula, Represents x i A power point within the k-distance neighborhood; Represents x i and The Euclidean distance between .
[0069] Then calculate the local reachable density of each power point, and the mathematical formula is expressed as:
[0070]
[0071] Finally, the outlier factor of the power point is calculated, and the mathematical formula is expressed as:
[0072]
[0073] If the local outlier factor of a power point is close to 1, it means that the power point is slightly different from its neighborhood and is considered a normal point; if the local outlier factor of a power point is much greater than 1, it means that the power point is significantly different from its neighborhood and is considered an outlier. This algorithm is used to detect and eliminate outliers in photovoltaic power generation.
[0074] Step 1-2: Use the multi-layer perceptron regression algorithm to fill in the missing data. The non-missing and missing data of photovoltaic power generation are used as training sets and test sets respectively. The multi-layer perceptron regression algorithm is used to train in the training set, and then the trained model is used to predict the test set, and finally the prediction results are used to fill in the missing data. The multi-layer perceptron model consists of an input layer, a hidden layer, and an output layer. The mathematical formula for each node from the input layer to the hidden layer and then to the output layer is expressed as:
[0075] x ij =g(W i X i-1 +b i-1 ) (4)
[0076] Where x ij The value of the jth neuron in the i-th layer, W i represents the weight vector of the jth neuron in the i-1th layer to the i-th layer; X i-1 b represents the vector of all neuron values in the i-1th layer; i-1 represents the bias term of the i-1th layer; g represents the activation function, and the ReLU function is used as the activation function.
[0077] The loss function of the multilayer perceptron regression algorithm is expressed in mathematical formula as follows:
[0078]
[0079] In the formula, Represents the output value of the multilayer perceptron regression algorithm; y i Represents the true value, and the multilayer perceptron regression algorithm stops calculating when the loss function value converges.
[0080] Step 2: Feature selection. By selecting representative features from multiple features through feature selection and reducing irrelevant feature information, the difficulty of model learning can be reduced and the training efficiency can be improved. The present invention adopts the distance correlation coefficient method to determine the correlation between each feature and the power generation power, and selects the features with a large correlation with the power generation power. Taking instantaneous radiation, scattered radiation, normal radiation, air temperature, air pressure, humidity and backplane temperature as features, and photovoltaic power generation power as a label, the Pearson correlation coefficient and distance correlation coefficient of each feature and the label are calculated, and the mathematical formula is expressed as follows:
[0081]
[0082] In the formula, x represents a certain feature, y represents the power generation, and v 2 (x,x) represents the feature distance square; v 2 (y,y) represents the distance variance of the generated power; v 2 (x,y) represents the distance covariance between the feature and the generated power.
[0083] The distance correlation coefficient between each feature and photovoltaic power generation is obtained by calculation, such as Figure 1 As shown in the figure, the distance correlation coefficient between air pressure and photovoltaic power generation is the smallest, and the correlation is very weak; the distance correlation coefficients between other features and photovoltaic power generation are all high, and the correlation is very strong. Therefore, in order to improve the efficiency of the model during training, the feature of air pressure is removed.
[0084] Step 3: Train sub-models. Divide the data into training set, validation set and test set. Train the support vector machine regression (SVR), multivariate linear regression (MLR) and Bayesian ridge regression (BRR) models in the training set, and verify the trained models in the validation set. Select the best models as sub-models.
[0085] Support vector machine regression (SVR) is a machine learning method based on the risk minimization criterion to solve regression fitting problems. The principle of the algorithm is to perform linear regression in high-dimensional space to find an optimal hyperplane so that the total deviation of the sample points from the hyperplane is minimized. Its mathematical formula is expressed as:
[0086]
[0087] In the formula, ω represents the weight vector, represents the kernel function, and b represents the bias.
[0088] Multiple linear regression (MLR) is analyzed by establishing a mathematical model between the dependent variable and its multiple independent variables. The mathematical formula of the multiple linear regression model with independent variables is expressed as:
[0089] y=β0+β1x1+...+β m x m +ε (8)
[0090] Where β0,β1,...β m represents the regression coefficient; y represents the dependent variable; x0, x1, ...x m represents the independent variable; ε represents the random error.
[0091] Bayesian Ridge Regression (BRR) is a regression algorithm based on Bayesian theory. It introduces the L2 regularization term in the estimation process, so it is called Bayesian Ridge Regression. The mathematical formula of Bayesian linear regression is:
[0092]
[0093] In the formula, m represents the dimension of the sample space; ω represents the parameter vector; x represents the input vector; represents a nonlinear function of x.
[0094] Step 4: Multi-data model fusion prediction. Use the best sub-models obtained through training to predict the data in the test set, and use the Q-learning algorithm to fuse the prediction results of each sub-model. The process of the Q-learning algorithm is as follows: Figure 2 The multi-data model fusion process is shown in Figure 3 As shown, it can be expressed as a Markov decision process {S, A, R}, where S represents the state set, A represents the action set, and R represents the reward function, where the state and action are a set of weights ω1, ω2, ..., ω n , record the state at time t Record the action at time t The reward function is the Fisher information, which can reflect the accuracy of parameter estimation. The larger the Fisher information, the higher the accuracy of parameter estimation. The calculation process of Fisher information is as follows.
[0095] First, we obtain the mathematical expression of the model prediction value:
[0096] M i =M0+ΔM i (10)
[0097] Where M i represents the prediction power of the ith model, M0 represents the true value, ΔM i represents the uncertainty of the power predicted by the ith model. Assume in represents the variance of the prediction power of the ith model.
[0098] The predicted power of the model conforms to the Gaussian distribution, that is:
[0099]
[0100] Since all models are independent, their joint probability density is:
[0101]
[0102] Where, Γ=(M1,M2,...M n ).
[0103] Finally, the Fisher information is calculated, and its mathematical expression is:
[0104]
[0105] In the fusion process, a state is obtained from the state set and an action is selected from the action set. After the action is taken, the fusion system gives a reward value according to the reward function and switches to the next state. The fusion system learns repeatedly until it reaches the terminal state and the Q value converges to the maximum value. Its update mechanism is expressed by the Bellman equation:
[0106]
[0107] In the formula, represents the Q value before the update at time t; represents the updated Q value at time t; Indicates the current status; Indicates the action taken in the current state; Indicates the next state; Indicates status The actions that can be taken under the condition of , δ represents the learning rate, which controls the ratio of new and old experiences in training; γ represents the discount factor, which indicates the importance of future rewards.
[0108] A system for implementing a photovoltaic power prediction method of multiple data model fusion of the present invention comprises a data preprocessing module, a feature selection module, a sub-model training module, and a multiple data model fusion prediction module connected in sequence, wherein the data preprocessing module, the feature selection module, the sub-model training module, and the multiple data model fusion prediction module correspond to the technical contents of steps 1 to 4 of the method of the present invention respectively.
[0109] like Figure 4 As shown, by comparing the results of a single model and other multi-data model fusion methods through four evaluation indicators, it can be seen that the error indicators of the multi-data model fusion prediction method proposed in the present invention are significantly reduced, and the prediction accuracy is higher.
[0110] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A photovoltaic power generation prediction method based on multi-data model fusion, comprising the following steps: Step 1: Data preprocessing: Step 1-1: Use the local outlier factor (LOF) algorithm to detect and remove outliers; The degree of outlier of each power point is evaluated by calculating the outlier factor of each power point. First, the Euclidean distance of each power point to other power points is calculated. Let k∈N+, and the power point is x i , power point x i and the kth point closest to it The distance between them is denoted as d k (x i ), then the power point x i The reachable distance to all power points within its k-distance neighborhood is: d r (x i ,x i o )=max{d k (x i o ),d(x i ,x i o )} (1) In the formula, Represents x i A power point within the k-distance neighborhood; Represents x i and The Euclidean distance between Then calculate the local reachable density of each power point, and the mathematical formula is expressed as: Finally, the outlier factor of the power point is calculated, and the mathematical formula is expressed as: If the local outlier factor of a power point is close to 1, it means that the power point is slightly different from its neighborhood and is considered a normal point; if the local outlier factor of a power point is much greater than 1, it means that the power point is significantly different from its neighborhood and is considered an outlier. This algorithm is used to detect and remove abnormal values in photovoltaic power generation. Step 1-2: Use the multi-layer perceptron regression algorithm to fill in the missing data; use the non-missing and missing data of photovoltaic power generation as the training set and test set respectively, use the multi-layer perceptron regression algorithm to train in the training set, then use the trained model to predict the test set, and finally use the prediction results to fill in the missing data; the multi-layer perceptron model consists of an input layer, a hidden layer and an output layer. The mathematical formula of each node from the input layer to the hidden layer and then to the output layer is expressed as: x ij =g(W i X i-1 +b i-1 ) (4) Where x ij The value of the jth neuron in the i-th layer, W i represents the weight vector of the jth neuron in the i-1th layer to the i-th layer; X i-1 b represents the vector of all neuron values in the i-1th layer; i-1 represents the bias term of the i-1th layer; g represents the activation function, and the ReLU function is used as the activation function; The loss function of the multilayer perceptron regression algorithm is expressed in mathematical formula as follows: In the formula, Represents the output value of the multilayer perceptron regression algorithm; y i Represents the true value, and the multilayer perceptron regression algorithm stops calculating when the loss function value converges; Step 2: Feature selection: Select representative features from multiple features through feature selection, reduce irrelevant feature information, reduce the difficulty of model learning, and improve training efficiency; use the distance correlation coefficient method to determine the correlation between each feature and power generation, and select features with high correlation with power generation; use instantaneous radiation, scattered radiation, normal radiation, air temperature, air pressure, humidity and backplane temperature as features, and photovoltaic power generation as labels, calculate the Pearson correlation coefficient and distance correlation coefficient of each feature and label, and the mathematical formula is as follows: In the formula, x represents a certain feature, y represents the power generation, and v 2 (x,x) represents the feature distance square; v 2 (y,y) represents the distance variance of the generated power; v 2 (x,y) represents the distance covariance between the feature and the generated power; The distance correlation coefficient D(x,y)∈[0,1] between each feature and photovoltaic power generation is obtained by calculation. The closer it is to 1, the stronger the correlation between the selected feature, label and power generation; the closer it is to 0, the weaker the correlation between the selected feature and power generation; Step 3: Train sub-models; divide the data into training set, validation set and test set, train support vector machine regression (SVR), multivariate linear regression (MLR) and Bayesian ridge regression (BRR) models in the training set, and verify the trained models in the validation set, and select the best models as sub-models; Support vector machine regression (SVR) is a machine learning method based on the risk minimization criterion to solve regression fitting problems. The principle of the algorithm is to perform linear regression in high-dimensional space to find an optimal hyperplane so that the total deviation of the sample points from the hyperplane is minimized. Its mathematical formula is expressed as: In the formula, ω represents the weight vector, represents the kernel function, b represents the bias; Multiple linear regression (MLR) is analyzed by establishing a mathematical model between the dependent variable and its multiple independent variables. The mathematical formula of the multiple linear regression model with independent variables is expressed as: y=β0+β1x1+...+β m x m +e (8) Where β0,β1,...β m represents the regression coefficient; y represents the dependent variable; x0, x1, ...x m represents the independent variable; ε represents the random error; Bayesian Ridge Regression (BRR) is a regression algorithm based on Bayesian theory. It introduces the L2 regularization term in the estimation process, so it is called Bayesian Ridge Regression. The mathematical formula of Bayesian linear regression is expressed as: In the formula, m represents the dimension of the sample space; ω represents the parameter vector; x represents the input vector; represents a nonlinear function of x; Step 4: Multi-data model fusion prediction; Use the best sub-models obtained through training to predict the data in the test set, and use the Q-learning algorithm to fuse the prediction results of each sub-model; the fusion process can be expressed as a Markov decision process {S, A, R}, where S represents the state set, A represents the action set, and R represents the reward function, where the state and action are a set of weights ω1, ω2, ..., ω n , record the state ω at time t s1 ,ω s2 ,...,ω sn for Remember the action at time t a1 ,ω a2 ,...,ω an for The reward function is Fisher information, which can reflect the accuracy of parameter estimation. The larger the Fisher information, the higher the accuracy of parameter estimation. The calculation process of Fisher information is as follows; First, we obtain the mathematical expression of the model prediction value: M i =M0+ΔM i (10) Where M i represents the prediction power of the i-th model, M0 represents the true value, ΔM i represents the uncertainty of the power predicted by the i-th model; assuming in represents the variance of the prediction power of the ith model; The predicted power of the model conforms to the Gaussian distribution, that is: Since all models are independent, their joint probability density is: Where, Γ=(M1,M2,...M n ); Finally, the Fisher information is calculated, and its mathematical expression is: In the fusion process, a state is obtained from the state set and an action is selected from the action set. After the action is taken, the fusion system gives a reward value according to the reward function and switches to the next state. The fusion system learns repeatedly until it reaches the terminal state and the Q value converges to the maximum value. Its update mechanism is expressed by the Bellman equation: In the formula, represents the Q value before the update at time t; represents the updated Q value at time t; Indicates the current status; Indicates the action taken in the current state; Indicates the next state; Indicates status The actions that can be taken under the condition of , δ represents the learning rate, which controls the ratio of new and old experiences in training; γ represents the discount factor, which indicates the importance of future rewards.
2. A system for implementing a photovoltaic power prediction method based on multi-data model fusion according to claim 1, characterized in that: It includes a data preprocessing module, a feature selection module, a sub-model training module, and a multi-data model fusion prediction module connected in sequence, among which: The data preprocessing module uses the local outlier factor (LOF) algorithm to detect and remove outliers, and the multi-layer perceptron regression algorithm to fill in missing data; The feature selection module selects representative features from multiple features through feature selection, reduces irrelevant feature information, reduces the difficulty of model learning, and improves training efficiency; The sub-model training module divides the data into training set, validation set and test set. Support vector machine regression (SVR), multivariate linear regression (MLR) and Bayesian ridge regression (BRR) models are trained in the training set, and the trained models are verified in the validation set. The best models are selected as sub-models. The multi-data model fusion prediction model uses the best sub-models obtained through training to predict the data in the test set, and uses the Q-learning algorithm to fuse the prediction results of each sub-model.
