Multi-model integrated distributed photovoltaic short-term power prediction method and system
In distributed photovoltaic short-term power prediction, using multi-model integration method to find similar prediction days from historical data and optimize model weights, the problem of inconsistent prediction effects of existing models is solved, and higher accuracy and stability are achieved.
Patent Information
- Application Number
- CN202411041526.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-19
AI Technical Summary
In the field of distributed short-term photovoltaic power prediction, the prediction effects of existing models are uneven and there is a risk of prediction failure. How to improve the stability and accuracy of the model is an important issue.
Using multi-model integration method, we can find similar prediction days from historical data, and optimize and combine multiple prediction models based on similar prediction days, build the best prediction model dynamically, and optimize the model weights using random convergence to obtain the prediction days. Best prediction results.
Through the multi-model integration method, the accuracy and stability of photovoltaic short-term power prediction are improved, the model is accurately predicted in different meteorological forecast scenarios, and the generalization ability and reference weight mediation between the models are enhanced.
Smart Images

Figure CN118983780A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of power systems, and more specifically, to a multi-model integrated distributed photovoltaic short-term power prediction method and system. Background Art
[0002] In recent years, photovoltaic power prediction technology has gradually developed and matured, and photovoltaic power prediction systems have also been put into practical use, playing an important role in formulating day-ahead power generation plans, promoting the consumption of new energy, and ensuring the economic and safe operation of power grids.
[0003] Photovoltaic power forecasting can be divided into ultra-short-term, short-term, and medium- and long-term forecasts according to the time scale. Among them, short-term power forecasting refers to the power forecast for the next 1 to 3 days. In the field of distributed short-term photovoltaic power forecasting, there are currently a large number of physical models, statistical models, and machine learning models in this field. Different models have different adaptability in different forecasting scenarios, and the forecasting effects are uneven. Regardless of which forecasting model is used, there is a risk of forecast failure. How to improve the stability of the forecasting model is an important issue that affects the application effect of the photovoltaic power forecasting system.
[0004] In view of the above problems, a multi-model integrated distributed photovoltaic short-term power prediction method and system is urgently needed. Summary of the invention
[0005] In order to address the deficiencies in the prior art, the present invention provides a multi-model integrated distributed photovoltaic short-term power prediction method and system. The method finds similar prediction days from historical data, and optimizes and combines multiple prediction models based on similar prediction days to dynamically obtain the best prediction result within the prediction day.
[0006] The present invention adopts the following technical solution.
[0007] The first aspect of the present invention relates to a multi-model integrated distributed photovoltaic short-term power prediction method, the method comprising the following steps: collecting and calculating the weather forecast for the prediction day, and searching for similar historical prediction days from a pre-generated data set; respectively obtaining the average prediction accuracy of multiple pre-trained prediction models for similar historical prediction days, and screening out multiple preferred prediction models based on the ranking of the average prediction accuracy; with the goal of minimizing the prediction error, constructing a power prediction model using multiple preferred prediction models, and optimizing the weight of each preferred prediction model in the power prediction model using a random convergence method, thereby obtaining a power prediction model for the prediction day; using the power prediction model for the prediction day, and using the weather forecast for the prediction day to predict the distributed photovoltaic short-term power on the prediction day.
[0008] Preferably, the weather forecast for the predicted day is collected and calculated, and similar historical predicted days are sought from a pre-generated data set, including: the pre-generated data set includes the total radiation peak, radiation direct-scatter ratio, maximum temperature, and minimum temperature on the predicted day; wherein, the similar historical predicted day is extracted from all historical days in the pre-generated data set based on the Euclidean distance between the data vector of the predicted day and the data vector in the pre-generated data set.
[0009] Preferably, the average prediction accuracy of multiple pre-trained prediction models for similar historical prediction days is obtained respectively, including: multiple pre-trained prediction models are obtained after training based on pre-generated data sets; the pre-trained prediction models at least include a multivariate linear regression model, a polynomial regression model, a support vector machine model, a BP neural network model, and a RBF neural network model.
[0010] Preferably, multiple preferred prediction models are selected based on the ranking of average prediction accuracy, including: calculating the average prediction accuracy of each pre-trained prediction model for historically similar prediction days; if the average prediction accuracy of any model is greater than or equal to a preset accuracy threshold, any model is taken as a preferred prediction model; and traversing all pre-trained prediction models to obtain multiple preferred prediction models.
[0011] Preferably, with the goal of minimizing the prediction error, a power prediction model is constructed using multiple preferred prediction models, including: the power prediction model is:
[0012]
[0013] In the formula, n is the model number of multiple preferred prediction models, N is the number of models of multiple preferred prediction models, β n is the weight of the nth preferred prediction model, P n is the prediction power of the nth preferred prediction model.
[0014] Preferably, the calculation process of the prediction power of the nth preferred prediction model is: training the nth prediction model through a pre-generated data set to obtain the nth preferred prediction model; combining multiple historical similar prediction days and prediction days into a similar data set, and inputting the similar data set into the nth preferred prediction model to obtain the prediction power of the nth preferred prediction model.
[0015] Preferably, a random convergence method is used to optimize the weight of each preferred prediction model in the power prediction model, including: randomly setting multiple N-dimensional weight vectors, substituting the multiple N-dimensional weight vectors into the power prediction model respectively, and solving the power prediction model to obtain multiple predicted powers; using the predicted power and the actual power to calculate the model deviation, and taking the N-dimensional weight vector corresponding to the predicted power with the smallest deviation as the optimal weight vector; based on the optimal weight vector, regenerating multiple N-dimensional weight vectors around the optimal weight vector, and repeating the above steps until the termination condition is met, and taking the best solution of the last iteration as the weight of each preferred prediction model in the power prediction model.
[0016] Preferably, the N-dimensional weight vector at the t+1th iteration is:
[0017]
[0018] Where, X rand is the random individual position, X(t) is the N-dimensional weight vector at the tth iteration, X * (t) is the best solution at the tth iteration,
[0019] A=2a*r1-a,C=2*r2,r1, r2 and p are random numbers between [0,1],a is a real number between [0,2] that decreases with the increase of the number of iterations t,
[0020] b is a constant, and l is a random number between [-1,1].
[0021] Preferably, the power prediction model of the prediction day is adopted, and the distributed photovoltaic short-term power of the prediction day is predicted using the weather forecast of the prediction day, including: updating the power prediction model using the optimal solution obtained by random convergence method; inputting the weather forecast of the prediction day into the updated power prediction model to obtain the distributed photovoltaic short-term power prediction value of the prediction day.
[0022] The second aspect of the present invention relates to a multi-model integrated distributed photovoltaic short-term power prediction system using the steps of the method in the first aspect of the present invention, the system including an acquisition module, a screening module, an optimization module and a prediction module; wherein the acquisition module is used to collect and calculate the weather forecast for the prediction day, and find similar historical prediction days from a pre-generated data set; the screening module is used to obtain the average prediction accuracy of multiple pre-trained prediction models for similar historical prediction days, and select multiple preferred prediction models based on the ranking of the average prediction accuracy; the optimization module is used to construct a power prediction model using multiple preferred prediction models with the goal of minimizing the prediction error, and optimize the weight of each preferred prediction model in the power prediction model using a random convergence method, so as to obtain the power prediction model for the prediction day; the prediction module is used to use the power prediction model of the prediction day to predict the distributed photovoltaic short-term power of the prediction day using the weather forecast of the prediction day.
[0023] The beneficial effect of the present invention is that, compared with the prior art, the method and system for distributed photovoltaic short-term power prediction with multi-model integration in the present invention dynamically obtain the best prediction result within the prediction day by finding similar prediction days from historical data and optimizing and combining multiple prediction models based on similar prediction days. The present invention is effective and reliable. On the one hand, it improves the generalization ability of the model prediction process by pre-training multiple prediction models. On the other hand, it selects some models by similar days and adjusts the reference weights between models to accurately predict the same period through typical photovoltaic power generation behavior, thereby ensuring the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 A schematic diagram of the steps of a multi-model integrated distributed photovoltaic short-term power prediction method of the present invention;
[0025] Figure 2 A schematic diagram of the relationship between predicted power and actual power in a distributed photovoltaic short-term power prediction method with multi-model integration according to the present invention;
[0026] Figure 3 The present invention is a schematic diagram of the module structure of a multi-model integrated distributed photovoltaic short-term power prediction system. DETAILED DESCRIPTION
[0027] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in the present invention are only embodiments of a part of the present invention, not all embodiments. Based on the spirit of the present invention, all other embodiments not recorded in the present invention obtained by ordinary technicians in this field according to the embodiments recorded in the present invention without making creative work should belong to the protection scope of the present invention.
[0028] Figure 1 The following is a schematic diagram of the steps of a multi-model integrated distributed photovoltaic short-term power prediction method of the present invention. Figure 1 As shown, the first aspect of the present invention relates to a multi-model integrated distributed photovoltaic short-term power prediction method, the method comprising steps 1 to 4.
[0029] Step 1: Collect and calculate the weather forecast for the forecast day, and find similar historical forecast days from the pre-generated dataset.
[0030] As mentioned above, in actual operation, the dispatcher usually builds multiple models based on the same training set. The model parameters are often controlled by the model builder, and the decision-maker usually does not intervene in the model construction process. The parallel combination prediction model involves the decomposition of the training set and is not suitable for multi-model decision-making scenarios. In the prior art, the Bagging framework and multi-model weighted prediction method are often used to implement integrated learning between multiple models.
[0031] Among them, the Bagging framework reduces the total error by reducing the variance of weak learners, constructs multiple training sample sets through extraction with replacement, and then constructs multiple sub-learning models at the same time. For regression problems, the average method is used to add the results of the sub-models to obtain the final prediction result. Multi-model weighted prediction also predicts by constructing multiple models. The difference is that each model uses the same training set, and the final prediction result is obtained by weighting each model in a weighted manner, replacing the simple average method in the Bagging framework. However, the existing parallel combination prediction model lacks consideration of the applicability issues in different weather forecast scenarios.
[0032] In order to solve the optimization problem of multi-model combination application and improve the accuracy and stability of the prediction system, the present invention extracts similar day data of the day to be predicted in advance, and uses the correlation between the weather forecast and power of similar days as a reference to implement the prediction of the prediction day. By selecting similar forecast days, the prediction accuracy of multiple models under similar meteorological forecast conditions is evaluated, and the prediction model is optimized; then, the optimal combination model is constructed using a random sampling optimization algorithm to achieve the best prediction results under the meteorological forecast conditions. Using this method for daily rolling judgment can achieve the best results in different meteorological forecast scenarios.
[0033] First, historical data of distributed photovoltaics are collected to build a historical data set. In this process, the collected data content may include weather forecast information for different days, such as historical numerical weather forecast data, historical active power data, and historical meteorological resource monitoring data. These data are calculated to obtain the data features in the data set.
[0034] Preferably, the weather forecast for the predicted day is collected and calculated, and similar historical predicted days are sought from a pre-generated data set, including: the pre-generated data set includes the total radiation peak value, radiation direct-scatter ratio, maximum temperature, and minimum temperature on the predicted day; wherein, the similar historical predicted days are extracted from all historical days in the pre-generated data set based on the Euclidean distance between the data vector of the predicted day and the data vector in the pre-generated data set.
[0035] In the present invention, the daily characteristic vector of numerical weather forecast is first constructed, and the characteristic values include total radiation peak value, radiation direct diffuse ratio, maximum temperature, minimum temperature and other indicators. Among them, the calculation formula of radiation direct diffuse ratio is:
[0036]
[0037] In the formula, DR t and DN t is the direct radiation and scattered radiation at the measuring time t, and Y0 is the sum of all measuring periods in a day.
[0038] The total radiation peak is obtained by calculating the data at multiple measurement times in a day, thereby comparing and analyzing all the collected data.
[0039] Each indicator uses the Min-Max method to achieve data normalization, and the Euclidean distance calculation is used to characterize the similarity between the historical daily numerical weather forecast results and the day to be predicted. After obtaining the similarity of the data, the similarity is sorted according to the Euclidean distance, so that multiple historical days that meet the conditions are selected as similar day data, and the prediction data set is constructed together with the current prediction day.
[0040] The calculation formula of Euclidean distance is:
[0041]
[0042] In the formula, x ik is the data of the kth dimension in the historical day, x 0k is the data of the kth dimension in the prediction day. After all dimensions of data participate in the Euclidean distance calculation, the calculation dimensions of two vectors are obtained.
[0043] By the above method, the Euclidean distance between each different historical day and the day to be predicted is analyzed, and multiple historical days with the smallest Euclidean distance values are taken as similar days. In one embodiment, all Euclidean distance calculation results are sorted by similarity, a threshold k% is set, and historical days with a similarity ranking less than k% are taken as similar days.
[0044] Step 2, respectively obtain the average prediction accuracy of multiple pre-trained prediction models for similar historical prediction days, and select multiple preferred prediction models based on the ranking of the average prediction accuracy.
[0045] Preferably, the average prediction accuracy of multiple pre-trained prediction models for historical similar prediction days is obtained respectively, including: multiple pre-trained prediction models are obtained after training based on the pre-generated data set; the pre-trained prediction models at least include a multivariate linear regression model, a polynomial regression model, a support vector machine model, a BP neural network model, and a RBF neural network model.
[0046] In the above model, the multivariate linear regression model uses the data dimensions mentioned above, such as irradiance and temperature, as input factors and output power as the prediction result. The calculation formula is:
[0047] h θ (x) = θ0 + θ1x1 + θ2x2 + ... + θ N x N
[0048] θ0 to θ N are the training parameters of the regression model, x1 to x N is the input data of the regression model, that is, the N-dimensional vector mentioned above.
[0049] During the pre-training process, the above model uses the data of all historical days in the historical data set as input conditions for training. This training result is not only implemented for the forecast day or similar days, so the calculation results of the model can be widely adapted to various types of forecast days. At the same time, due to the characteristics of the model itself, the forecast results also contain systematic errors that are difficult to avoid in the model itself. By implementing multi-model fitting and adjusting the weights, the systematic errors between multiple models can be offset in a reasonable way on a certain forecast day.
[0050] The polynomial regression model is also trained in a similar way. The data conditions of the polynomial model are similar to the previous model, and the output power is:
[0051]
[0052] In the formula, a0 to a N Train the parameters for the model.
[0053] In addition, the power calculation process of the support vector machine model is:
[0054]
[0055] In the formula, The input vector x i The corresponding dual weight of the support weight,
[0056] α i The input vector x i The corresponding support weight,
[0057] k(·) is the kernel function of the support vector machine,
[0058] b is the intercept.
[0059] In the above model, the input vector x i The corresponding support weights are parameters for model training.
[0060] In a similar manner, the present invention also supports the use of BP neural network, RBF neural network and other models in a similar manner to input the radiation, temperature, humidity and other factors of numerical weather forecast, and output power prediction results. Since there are many types of prediction models and the selection of model parameters is involved, the above is a model example, and the method of the present invention is applicable to any model.
[0061] Use the pre-trained M prediction models to predict each day in the similar forecast day set, calculate the average prediction accuracy of each model for the data set, and select the top N models with higher prediction accuracy to participate in subsequent integrated calculations.
[0062] Preferably, multiple preferred prediction models are selected based on the ranking of average prediction accuracy, including: calculating the average prediction accuracy of each pre-trained prediction model for historically similar prediction days; if the average prediction accuracy of any model is greater than or equal to a preset accuracy threshold, any model is taken as a preferred prediction model; and traversing all pre-trained prediction models to obtain multiple preferred prediction models.
[0063] The above-mentioned calculation method for the average prediction accuracy can adopt commonly used calculation formulas, such as MSE, etc.
[0064] Step 3, with the goal of minimizing the prediction error, a power prediction model is constructed using multiple preferred prediction models, and the weight of each preferred prediction model in the power prediction model is optimized using a random convergence method to obtain a power prediction model for the prediction day.
[0065] Preferably, a combined forecast model is established with the minimum forecast error under similar forecast days as the optimization goal, and the obtained combined model is a power prediction model, and the calculation formula is as follows:
[0066]
[0067] In the formula, n is the model number of multiple preferred prediction models, N is the number of models of multiple preferred prediction models, β n is the weight of the nth preferred prediction model, P n is the prediction power of the nth preferred prediction model.
[0068] Preferably, the calculation process of the prediction power of the nth preferred prediction model is: training the nth prediction model through a pre-generated data set to obtain the nth preferred prediction model; combining multiple historical similar prediction days and prediction days into a similar data set, and inputting the similar data set into the nth preferred prediction model to obtain the prediction power of the nth preferred prediction model.
[0069] In the process of calculating the combined weight, the predicted power pre-calculated by each preferred model is input into the above formula, so as to solve and analyze the weight of each iteration and the accuracy of the combined solution of multiple preferred models.
[0070] The process of training to obtain the optimal combination function of N models is implemented iteratively. First, a set of weights of each prediction model is used as a sample, and the initial position of the model is obtained by random sampling. Then, the optimal solution is obtained by gradually narrowing the range through a spiral. In this process, a small-scale logarithmic spiral equation is also used to obtain the random quantity in the spatial position to prevent the occurrence of local optimal problems.
[0071] Preferably, a random convergence method is used to optimize the weight of each preferred prediction model in the power prediction model, including: randomly setting multiple N-dimensional weight vectors, substituting the multiple N-dimensional weight vectors into the power prediction model respectively, and solving the power prediction model to obtain multiple predicted powers; using the predicted power and the actual power to calculate the model deviation, and taking the N-dimensional weight vector corresponding to the predicted power with the smallest deviation as the optimal weight vector; based on the optimal weight vector, regenerating multiple N-dimensional weight vectors around the optimal weight vector, and repeating the above steps until the termination condition is met, and taking the best solution of the last iteration as the weight of each preferred prediction model in the power prediction model.
[0072] In the above formula, multiple N-dimensional weight vectors are randomly set, and then the optimal weight vector is found in the power prediction model through the multiple vectors. The search process is implemented based on the model deviation. The model deviation can be obtained based on the difference between the predicted power calculated by the combined model and the actual power.
[0073] Preferably, the N-dimensional weight vector at the t+1th iteration is:
[0074]
[0075] Where, X rand is the random individual position, X(t) is the N-dimensional weight vector at the tth iteration, X * (t) is the best solution at the tth iteration,
[0076] A=2a*r1-a,C=2*r2,r1, r2 and [are random numbers between [0,1],a is a real number between [0,2] that decreases as the number of iterations t increases,
[0077] b is a constant, and l is a random number between [-1,1].
[0078] By iterating step by step in the above manner, the final optimal weight is obtained. The optimal weight is substituted into the formula to implement the calculation of the power prediction model.
[0079] Step 4: Use the power prediction model of the prediction day and the weather forecast of the prediction day to predict the distributed photovoltaic short-term power of the prediction day.
[0080] Preferably, a power prediction model for the prediction day is adopted, and the distributed photovoltaic short-term power for the prediction day is predicted using the weather forecast for the prediction day, including: updating the power prediction model using the optimal solution obtained by a random convergence method; inputting the weather forecast for the prediction day into the updated power prediction model to obtain the distributed photovoltaic short-term power prediction value for the prediction day.
[0081] In the present invention, an optimal model is obtained and the data of the prediction day is substituted into the model, thereby obtaining an accurate distributed photovoltaic short-term power prediction result.
[0082] The similar day selection scheme proposed in the present invention involves the daily characteristic vector of the numerical weather forecast. The characteristic values include indicators such as total radiation peak value, radiation direct-scatter ratio, maximum temperature, minimum temperature, etc., which can better distinguish different weather processes and achieve the best similar day matching.
[0083] The present invention proposes a distributed photovoltaic short-term power prediction method with multi-model integration, which achieves the best prediction results under different weather forecast scenarios by constructing the optimal model under similar days. The method can make multi-model decisions without interfering with the training of a single model, and achieve the optimal model under the weather forecast conditions, thereby improving the accuracy and stability of the prediction system at the model level.
[0084] Figure 2 The figure is a schematic diagram showing the relationship between the predicted power and the actual power in a distributed photovoltaic short-term power prediction method with multi-model integration according to the present invention. Figure 2As shown, the independent prediction results of the forecast day of the six models are used as input and substituted into the combined prediction model to achieve the prediction results under the multi-model integration. The RMSE value of the specific embodiment is 91.18%.
[0085] Figure 3 The schematic diagram of the module structure of a distributed photovoltaic short-term power prediction system with multi-model integration in the present invention is shown in FIG. Figure 3 As shown, the second aspect of the present invention relates to a distributed photovoltaic short-term power prediction system integrating multiple models using the method in the first aspect of the present invention, the system comprising an acquisition module, a screening module, an optimization module and a prediction module; wherein the acquisition module is used to collect and calculate the weather forecast for the prediction day, and to find similar historical prediction days from a pre-generated data set; the screening module is used to respectively obtain the average prediction accuracy of multiple pre-trained prediction models for similar historical prediction days, and to select multiple preferred prediction models based on the ranking of the average prediction accuracy; the optimization module is used to construct a power prediction model using multiple preferred prediction models with the goal of minimizing the prediction error, and to optimize the weight of each preferred prediction model in the power prediction model using a random convergence method, so as to obtain a power prediction model for the prediction day; the prediction module is used to use the power prediction model of the prediction day to predict the distributed photovoltaic short-term power of the prediction day using the weather forecast of the prediction day.
[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A multi-model integrated distributed photovoltaic short-term power prediction method, It is characterized in that The method comprises the following steps: Collect and calculate the weather forecast for the forecast day, and find similar historical forecast days from the pre-generated data set; Respectively obtaining average prediction accuracy rates of multiple pre-trained prediction models for the historical similar prediction days, and selecting multiple preferred prediction models based on the ranking of the average prediction accuracy rates; With the goal of minimizing the prediction error, a power prediction model is constructed using the multiple preferred prediction models, and the weight of each of the preferred prediction models in the power prediction model is optimized using a random convergence method, so as to obtain the power prediction model for the prediction day; The power prediction model of the prediction day is adopted, and the weather forecast of the prediction day is used to predict the distributed photovoltaic short-term power of the prediction day.
2. According to the multi-model integrated distributed photovoltaic short-term power prediction method of claim 1, Features: The collecting and calculating of the weather forecast for the forecast day and searching for similar historical forecast days from the pre-generated data set include: The pre-generated data set includes the total radiation peak value, radiation direct-scatter ratio, maximum temperature, and minimum temperature on the forecast day; wherein, The historical similar prediction day is extracted from all historical days of the pre-generated data set based on the Euclidean distance between the data vector of the prediction day and the data vector in the pre-generated data set.
3. According to the multi-model integrated distributed photovoltaic short-term power prediction method of claim 2, Features: The obtaining of the average prediction accuracy of the multiple pre-trained prediction models for the historical similar prediction days includes: The plurality of pre-trained prediction models are obtained after training based on the pre-generated data set; The pre-trained prediction model at least includes a multivariate linear regression model, a polynomial regression model, a support vector machine model, a BP neural network model, and a RBF neural network model.
4. According to claim 3, a distributed photovoltaic short-term power prediction method based on multi-model integration, Features: The sorting based on the average prediction accuracy selects a plurality of preferred prediction models, including: Calculating the average prediction accuracy of each of the pre-trained prediction models for the historical similar prediction days; If the average prediction accuracy of any model is greater than or equal to a preset accuracy threshold, then the model is used as a preferred prediction model; All of the pre-trained prediction models are traversed to obtain a plurality of the preferred prediction models.
5. A distributed photovoltaic short-term power prediction method based on multi-model integration according to claim 4, Features: The method of constructing a power prediction model by using the plurality of preferred prediction models with the goal of minimizing the prediction error includes: The power prediction model is: Wherein, n is the model number of the plurality of preferred prediction models, N is the model number of the plurality of preferred prediction models, β n is the weight of the nth preferred prediction model, P n is the prediction power of the nth preferred prediction model.
6. A distributed photovoltaic short-term power prediction method based on multi-model integration according to claim 5, Features: The calculation process of the predicted power of the nth preferred prediction model is: Training the nth prediction model using the pre-generated data set to obtain the nth preferred prediction model; A plurality of historical similar forecast days and forecast days are combined into a similar data set, and the similar data set is input into the nth preferred forecast model to obtain the forecast power of the nth preferred forecast model.
7. A distributed photovoltaic short-term power prediction method based on multi-model integration according to claim 6, Features: The step of optimizing the weight of each of the preferred prediction models in the power prediction model by adopting a random convergence method includes: Randomly setting a plurality of N-dimensional weight vectors, respectively substituting the plurality of N-dimensional weight vectors into the power prediction model, and solving the power prediction model to obtain a plurality of predicted powers; Calculate the model deviation using the predicted power and the real power, and take the N-dimensional weight vector corresponding to the predicted power with the smallest deviation as the optimal weight vector; Based on the optimal weight vector, multiple N-dimensional weight vectors are regenerated around the optimal weight vector, and the above steps are repeated until the termination condition is met, and the best solution of the last iteration is used as the weight of each of the preferred prediction models in the power prediction model.
8. A distributed photovoltaic short-term power prediction method based on multi-model integration according to claim 7, Features: The N-dimensional weight vector at the t+1th iteration is: Where, X rand is the random individual position, X(t) is the N-dimensional weight vector at the tth iteration, X * (t) is the best solution at the tth iteration, A=2a*r 1 -a, C = 2*r 2 , r 1 、r 2 p is a random number between [0,1], a is a real number between [0,2] that decreases as the number of iterations t increases, n is a constant, and l is a random number between [-1,1].
9. A distributed photovoltaic short-term power prediction method based on multi-model integration according to claim 8, Features: The adopting the power prediction model of the prediction day and using the weather forecast of the prediction day to predict the distributed photovoltaic short-term power of the prediction day includes: Using the optimal solution obtained by the random convergence method, updating the power prediction model; The weather forecast for the forecast day is input into the updated power forecast model to obtain the distributed photovoltaic short-term power forecast value for the forecast day.
10. A distributed photovoltaic short-term power prediction system using a multi-model integrated method according to any one of claims 1 to 9, Features: The system includes a collection module, a screening module, an optimization module and a prediction module; wherein, The acquisition module is used to collect and calculate the weather forecast for the forecast day, and to find similar historical forecast days from the pre-generated data set; The screening module is used to respectively obtain the average prediction accuracy of multiple pre-trained prediction models for the historical similar prediction days, and screen out multiple preferred prediction models based on the ranking of the average prediction accuracy; The optimization module is used to construct a power prediction model using the multiple preferred prediction models with the goal of minimizing the prediction error, and optimize the weight of each of the preferred prediction models in the power prediction model using a random convergence method, so as to obtain the power prediction model for the prediction day; The prediction module is used to adopt the power prediction model of the prediction day and use the weather forecast of the prediction day to predict the distributed photovoltaic short-term power of the prediction day.
