A Short-Term Electric Load Forecasting Method and Device Based on Multi-Model Fusion

Abstract

The present invention discloses a short-term electric load forecasting method and device based on multi-model fusion. The method includes: respectively collecting electric load data and meteorological data to obtain original training data; constructing a hierarchical forecasting large model, where the hierarchical forecasting large model includes a first layer, a second layer, and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; at least one sub-model is included in each layer; training and optimizing the parameters of each layer in the hierarchical forecasting large model based on the original training data; and collecting electric load data and meteorological data in real time and inputting them into the trained hierarchical forecasting large model to forecast the short-term electric load. By constructing the hierarchical forecasting large model, the present invention fuses real-time data and the forecasting results of each sub-model in each layer of the hierarchical forecasting large model, thereby improving the accuracy and stability of the forecasting results.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a short-term power load forecasting method, device, electronic device, and computer-readable storage medium based on multi-model fusion. Background Art

[0002] Power load forecasting has the following important functions in the power system:

[0003] First, it can optimize resource allocation and power generation plans. For example, by predicting future power demands, the power system can formulate power generation plans in advance, reasonably arrange the operation mode of the power grid and the unit maintenance plan, thereby saving resources such as coal and oil, reducing power generation costs and operation costs, optimizing resource allocation, and further formulating a reasonable power source construction plan.

[0004] Second, it can ensure the safe and economic operation of the power system: The accuracy and stability of load forecasting are crucial for the safe and economic operation of the power system. It helps to avoid insufficient or excessive power supply, ensure the stable operation of the power system, reduce the risk of power outages, and improve the economic and social benefits of the power system.

[0005] Third, it provides a basis for the decision-making of power selling companies: In the spot market, signing strategies, trading strategies, etc. all rely on the results of load forecasting. Without accurate load forecasting, power selling companies cannot quote prices in the spot market, which may lead to high deviation fees.

[0006] In summary, power load forecasting is crucial for the stable operation, resource optimization, and economic benefits of the power system.

[0007] The existing methods for power load forecasting mainly include statistical forecasting methods and artificial intelligence methods based on machine models. Among them: Statistical forecasting methods mainly rely on statistical models, time series analysis, etc. Not only is the workload for implementation large, but it is often difficult to process non-linear and non-stationary power load data, and the ability to capture complex seasonal, trend, and periodic characteristics is limited, and the forecasting accuracy is difficult to guarantee. As a research hotspot, the widely used machine models in artificial intelligence methods mainly include backpropagation neural network, support vector machine, convolutional neural network (CNN), long short-term memory network (LSTM), stacked denoising autoencoder network (SDAE), and deep belief network (DBN), etc. Among them: As an emerging research direction in the field of artificial intelligence, deep learning has currently been applied in the field of power load forecasting, but there is still room for improvement in the forecasting effect.

[0008] Therefore, how to achieve accurate and stable forecasting of power load data has become an urgent problem to be solved currently. Summary of the Invention

[0009] In view of this, the main object of the present invention is to provide a short-term electric load forecasting method, device, electronic device and computer-readable storage medium based on multi-model fusion, so as to at least partially solve the above technical problems.

[0010] To achieve the above object, as the first aspect of the present invention, there is provided a short-term electric load forecasting method based on multi-model fusion, including the following steps:

[0011] Collect electric load data and meteorological data respectively to obtain original training data;

[0012] Construct a hierarchical prediction large model, which includes a first layer, a second layer and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; at least one sub-model is included in each layer;

[0013] Train and optimize the parameters of each layer in the hierarchical prediction large model based on the original training data;

[0014] Collect electric load data and meteorological data in real time and input them into the trained hierarchical prediction large model to predict short-term electric load.

[0015] According to a preferred embodiment of the present invention, the first layer includes a plurality of first sub-models connected in parallel, the second layer includes a plurality of second sub-models connected in parallel, and the third layer includes a third sub-model.

[0016] According to a preferred embodiment of the present invention, the types of the first sub-models are different, and at least include a KAN neural network model, and at least one model selected from a natural gradient boosting machine NGBoost, a multi-head quantum self-attention neural network prediction model MQSAPN, a regularized greedy forest RGF, a SLOTH model, and an iTransformer model; the types of the second sub-models are different, and at least two selected from a random forest model RF and a light gradient boosting machine lightGBM; the third sub-model is a partial least squares regression model PLSR.

[0017] According to a preferred embodiment of the present invention, the training and optimizing the parameters of each layer in the hierarchical prediction large model based on the original training data includes:

[0018] Initialize the parameter value ranges of each layer in the hierarchical prediction large model, and configure the initial optimization iteration times;

[0019] Randomly generate N populations, initialize the fitness of all individuals in the N populations, and configure the first optimization iteration times;

[0020] Perform the first optimization on N populations respectively based on the first optimized iteration times, and update the fitness to obtain the optimal individuals of the N populations;

[0021] Use the optimal individuals of the N populations as the initial population, and configure the second optimized iteration times to perform the second optimization to obtain the individual with the optimal fitness.

[0022] When the current iteration times meet the requirements of the second optimized iteration times and the total iteration times meet the requirements of the initial optimized iteration times, output the optimal individual as the parameters of each layer of the hierarchical prediction large model;

[0023] Where: N is a natural number.

[0024] According to a preferred embodiment of the present invention, each population contains m individuals, and the performing the first optimization on N populations respectively based on the first optimized iteration times, and updating the fitness to obtain the optimal individuals of the N populations includes:

[0025] Steps for updating and optimizing the m individuals of each population: successively perform triangular topology unit formation, global aggregation, and local aggregation on the m individuals to update the m individuals; calculate the fitness of the m individuals after the update, and update the optimal individual of this iteration according to the magnitude of the fitness;

[0026] Judge whether the current iteration times reach the first optimized iteration times;

[0027] If not satisfied, return to execute the steps for updating and optimizing the m individuals of each population;

[0028] If satisfied, compare the fitness of the optimal individuals obtained by each population in each layer iteration, and use the individual with the highest fitness as the optimal individual of the population;

[0029] Where: m is a natural number.

[0030] According to a preferred embodiment of the present invention, the second optimization is performed through the following formula:

[0031]

[0032] Where: is the position of the i-th individual at time t, f i is the fragrance value emitted by the i-th individual, p is the conversion probability, the optimal solution at time t is g * , Cauchy(0,1) and N(0,1) are the standard Cauchy distribution and the standard Gaussian distribution respectively, r is a random number in [0,1], T2 is the second optimized iteration times, and are the solutions corresponding to the positions of the j-th individual and the k-th individual at time t respectively, and both i and j are natural numbers less than or equal to N.

[0033] According to a preferred embodiment of the present invention, the second optimization further includes:

[0034] Randomly select 3 parent individuals in different dimensions;

[0035] Perform a vertical crossover operation on the d1-th, d2-th, and d3-th dimensions of the parent individuals to generate offspring individuals;

[0036] Compare the fitness values of the generated offspring individuals with those of the parent individuals, and retain the individuals with better fitness values as the individuals with the optimal fitness.

[0037] According to a preferred embodiment of the present invention, the offspring individuals are obtained through the following formula:

[0038] MS VC (i, d1) = n(qx(i, d1) + (1 - q)x(i, d2)) + (1 - n)x(i, d3)

[0039] Where: MS VC (i, d1) is the offspring individual obtained after the vertical operation, q ∈ [0, 1] and n ∈ [0, 1] are random numbers.

[0040] To solve the above technical problems, a short-term electric load forecasting device based on multi-model fusion is provided in the second aspect of the present invention, including:

[0041] An acquisition module for respectively acquiring electric load data and meteorological data to obtain original training data;

[0042] A construction module for constructing a hierarchical prediction large model, the hierarchical prediction large model including a first layer, a second layer, and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; at least one sub-model is included in each layer;

[0043] An optimization training module for training and optimizing the parameters of each layer in the hierarchical prediction large model based on the original training data;

[0044] A prediction module for real-time collecting electric load data and meteorological data and inputting them into the trained hierarchical prediction large model to predict short-term electric load.

[0045] To solve the above technical problems, an electronic device is provided in the third aspect of the present invention, including:

[0046] A processor; and

[0047] A memory storing computer-executable instructions, which when executed cause the processor to execute the method according to any one of the above.

[0048] To solve the above technical problems, a fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by a processor, the method described in any one of the above is implemented.

[0049] The short-term power load forecasting method, device, electronic device and computer storage medium based on multi-model fusion provided by the present invention construct, train and optimize a hierarchical prediction large model including a three-layer structure, wherein: the hierarchical prediction large model includes a first layer, a second layer and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; at least one sub-model is included in each layer. In this way, real-time data (load data and meteorological data) is processed by the models in the first layer to obtain a first prediction value; the first prediction value and the data to be predicted are processed by the models in the second layer to obtain a second prediction value, and the first prediction value and the second prediction value are processed by the models in the third layer to obtain the final power load forecasting result, so as to fuse the real-time data and the prediction results of each sub-model in each layer of the hierarchical prediction large model, and improve the accuracy and stability of the prediction results.

[0050] Compared with the prior art, it has at least the following beneficial effects:

[0051] 1. By constructing a hierarchical prediction large model, the real-time data and the prediction results of each sub-model in each layer of the hierarchical prediction large model are fused, and the accuracy and stability of the prediction results are improved.

[0052] 2. Each layer of sub-models is selected from different types, and a variety of tree ensemble models are fused with neural network models and quantum neural networks, effectively utilizing the prediction capabilities of a variety of models, and improving the accuracy and stability of the prediction results.

[0053] 3. By optimizing the parameters of each layer of the hierarchical prediction large model, the model prediction error is reduced, and the accuracy of the prediction results is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to make the technical problems solved by the present invention, the technical means adopted and the technical effects obtained more clear, the specific embodiments of the present invention will be described in detail below with reference to the drawings. However, it should be noted that the drawings described below are only the drawings of the exemplary embodiments of the present invention, and those skilled in the art can obtain the drawings of other embodiments without creative efforts based on these drawings.

[0055] Figure 1 It is a schematic diagram of the process of a short-term power load forecasting method based on multi-model fusion provided by an embodiment of the present invention;

[0056] Figure 2 It is a schematic structural diagram of the hierarchical prediction large model constructed in the embodiment of the present invention;

[0057] Figure 3 It is a schematic flowchart of optimizing the parameters of each layer of the hierarchical prediction large model in the embodiment of the present invention;

[0058] Figure 4 It is a schematic structural framework diagram of a short-term power load prediction device based on multi-model fusion provided by the embodiment of the present invention;

[0059] Figure 5 It is a block diagram of the structure of an exemplary embodiment of an electronic device according to the present invention;

[0060] Figure 6 It is a schematic diagram of an embodiment of a computer-readable medium of the present invention. Detailed implementation manners

[0061] Now, the exemplary embodiments of the present invention will be described more comprehensively with reference to the accompanying drawings. However, the exemplary embodiments can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, providing these exemplary embodiments enables the present invention to be more comprehensive and complete, and more convenient to fully convey the inventive concept to those skilled in the art. The same reference numerals in the figures denote the same or similar elements, components, or parts, and thus their repeated description will be omitted.

[0062] On the premise of conforming to the technical concept of the present invention, the features, structures, characteristics, or other details described in a specific embodiment do not exclude being combined in a suitable manner in one or more other embodiments.

[0063] In the description of the specific embodiments, the features, structures, characteristics, or other details described in the present invention are for enabling those skilled in the art to fully understand the embodiments. However, it does not exclude that those skilled in the art can practice the technical solutions of the present invention without one or more of the specific features, structures, characteristics, or other details.

[0064] The flowcharts shown in the accompanying drawings are only exemplary illustrations, and do not necessarily include all the contents and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined. Therefore, the actual execution order may change according to the actual situation.

[0065] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0066] It should be understood that although the terms "first", "second", "third", etc. may be used herein to describe various devices, elements, components or parts, this should not be limited by these terms. These terms are used to distinguish one from another. For example, a first device may also be called a second device without departing from the essential technical solution of the present invention.

[0067] The term "and / or" or "and / or" includes any one and all combinations of one or more of the associated listed items.

[0068] The embodiment of the present invention provides a short-term power load forecasting method based on multi-model fusion, such as Figure 1 As shown, the method includes:

[0069] S1, respectively collect power load data and meteorological data to obtain original training data;

[0070] Considering that the meteorological environment has a certain impact on the power load, for example, when the temperature is too low or too high, people will use electrical appliances such as air conditioners, which will obviously increase the power load. This application collects both power load data and meteorological data. Among them, meteorological data can include: temperature, humidity and air pressure.

[0071] Furthermore, in order to improve the prediction accuracy, this step can also pre-process the collected original training data, such as: missing value processing, normalization processing, etc. Among them: missing value processing can directly delete the data with missing values, or fill in the missing values ​​through mean interpolation, high-dimensional mapping, maximum likelihood estimation, matrix completion and other methods.

[0072] In addition, the original training data can be divided into training set and test set to facilitate subsequent training and testing.

[0073] S2. Build a large hierarchical prediction model.

[0074] The hierarchical prediction large model includes a first layer, a second layer, and a third layer connected in sequence. The input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer. Each layer includes at least one sub-model. In this way, the data to be processed is processed by the model in the first layer to obtain a first prediction value; the first prediction value and the data to be predicted are processed by the model in the second layer to obtain a second prediction value, and the first prediction value and the second prediction value are processed by the model in the third layer to obtain the final power load prediction result, thereby fusing the data to be processed and the prediction results of each sub-model in each layer of the hierarchical prediction large model to improve the accuracy and stability of the prediction result.

[0075] In the embodiment of the present invention, each layer of the hierarchical prediction large model may include only one sub-model or multiple sub-models. The types of each sub-model may be the same or different. If a layer contains multiple sub-models, the multiple sub-models may be connected in parallel or in series, and the present invention does not make specific limitations. When connected in parallel, the outputs of each sub-model in the same layer are used as the output results of this layer and can be input to the next layer or other layers.

[0076] In a preferred example, as Figure 2 shown, the hierarchical prediction large model includes a first layer, a second layer, and a third layer connected in sequence. The input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer. Each layer includes at least one sub-model. Among them: the first layer includes multiple first sub-models connected in parallel, the second layer includes multiple second sub-models connected in parallel, and the third layer includes a third sub-model.

[0077] Optionally, the types of each first sub-model are different, including: the KAN neural network model, and at least one model selected from the natural gradient boosting machine NGBoost, the multi-head quantum self-attention neural network prediction model MQSAPN, the regularized greedy forest RGF, the SLOTH model, and the iTransformer model; the types of the second sub-model are different, and at least two are selected from the random forest model RF and the light gradient boosting machine lightGBM; the third sub-model is a partial least squares regression model PLSR.

[0078] In a specific embodiment, as Figure 2 shown, the first layer includes: the NGBoost model, the MQSAPN model, the RGF model, the SLOTH model, the iTransformer model, and the KAN neural network model connected in parallel, the second layer includes: the RF model and the lightGBM model connected in parallel, and the third layer only includes one PLSR model.

[0079] In this embodiment, the collected load data and meteorological data are respectively output through the NGBoost model, MQSAPN model, RGF model, SLOTH model, iTransformer model, and KAN neural network model connected in parallel in the first layer to obtain six first prediction results for predicting the load data and meteorological data from different dimensions. Among them: The NGBoost model uses Gradient Boosting for prediction uncertainty estimation through probability prediction (including real-value output). By using the natural gradient, NGBoost overcomes the technical challenge of making general probability prediction difficult through gradient boosting. The MQSAPN model includes two parts: a multi-head quantum self-attention module and a variational quantum circuit prediction module. By performing quantum state encoding and the calculation of K, Q, and V on the input data at time steps respectively, and using the Gaussian function for the estimation method of self-attention coefficients, the data after quantum self-attention feature extraction is encoded into the variational prediction circuit again. After circuit evolution and measurement, the prediction result is finally obtained. The RGF model is a boosting algorithm based on decision trees, which improves the disadvantages of GBDT. The SLOTH model (Structured Learning and Task-based Optimization for Time Series Forecasting on Hierarchies) is a model for hierarchical time series forecasting, including two tree-based feature integration mechanisms, namely top-down convolution and bottom-up attention integration mechanisms, to utilize the information of the hierarchical structure to improve the prediction performance. The KAN neural network model is an improvement of the MLP, inspired by the Kolmogorov-Arnold representation theorem. Different from the MLP, KAN places the learnable activation function on the weights and has excellent learning ability. The iTransformer (Inverted Transformer) is an innovative time series prediction model, aiming to solve the problems of performance degradation and computational explosion of traditional Transformer models when predicting time series with a large range of backtracking windows.

[0080] Subsequently, the six first prediction results, as well as the collected load data and meteorological data, are respectively subjected to the first fusion process through the RF model and the lightGBM model connected in parallel in the second layer to obtain two second prediction results. Among them: The RF model is an ensemble learning method that improves the accuracy and stability of classification or regression by constructing multiple decision trees. The lightGBM model is based on the gradient boosting framework and has a faster training speed and lower memory consumption.

[0081] Finally, input the 6 first prediction results output by the first layer and the 2 second prediction results output by the second layer into the PLSR model of the third layer to obtain the final prediction result. Among them: The PLSR model finds the basic relationship between two matrices (X and Y), and finds the multi-dimensional direction in the X space to explain the multi-dimensional direction with the largest variance in the Y space, which is suitable when the prediction matrix has more variables than the observations and there is multicollinearity in the values of X.

[0082] S3. Train and optimize the parameters of each layer in the hierarchical prediction large model based on the original training data;

[0083] In the first optimization process, continuously search within the parameter range to obtain the optimal parameters. Among them, each triangular topological unit can represent a search individual, and different-sized similar triangular topological units are formed by aggregation as the basic evolutionary units, and the evolution of individuals within the unit is guided by the best individual in each triangular unit. This evolutionary population not only depends on the excellent individuals guided globally but also absorbs the effective positive information of the best individual in each unit, which helps to solve complex optimization problems and overcome the drawback of traditional methods falling into local extrema during global search.

[0084] Exemplarily, as Figure 3 shown, this step may include:

[0085] S301. Initialize the parameter value ranges of each layer in the hierarchical prediction large model and configure the initial optimization iteration times T0;

[0086] S302. Randomly generate N populations, initialize the fitness of all individuals in the N populations, and configure the first optimization iteration times T1;

[0087] An individual in each population is equivalent to a set of parameters (parameter 1, parameter 2,..., parameter d) of the hierarchical prediction large model. Among them: N is a natural number, and d is the total number of parameters of the hierarchical prediction large model.

[0088] Exemplarily, the prediction accuracy of the hierarchical prediction large model can be used as the individual fitness to initialize the fitness of all individuals in the N populations.

[0089] The first optimization iteration times T1 is the upper limit of the iteration times for each population.

[0090] S303. Perform the first optimization on the N populations respectively based on the first optimization iteration times, update the fitness, and obtain the optimal individuals of the N populations;

[0091] Taking each population containing m individuals as an example, where: m is a natural number. The following steps can be executed:

[0092] Step s11: Steps for updating and optimizing m individuals in each population: Sequentially perform triangular topological unit formation, global aggregation, and local aggregation on the m individuals to update the m individuals; calculate the fitness of the m individuals after the update, and update the optimal individual of this iteration according to the magnitude of the fitness.

[0093] Step s12: Determine whether the current iteration count has reached the first optimization iteration count T1.

[0094] Step s13: If not satisfied, return to execute Step s11.

[0095] Step s14: If satisfied, compare the fitness of the optimal individuals obtained from each layer of iteration of each population, and take the individual with the highest fitness as the optimal individual of that population; finally, N optimal individuals from N populations will be obtained.

[0096] S304: Use the optimal individuals of the N populations as the initial population, and configure the second optimization iteration count T2 to perform the second optimization to obtain the individual with the optimal fitness.

[0097] The second optimization is carried out through the following formula:

[0098]

[0099]

[0100] Where: is the position of the i-th individual at time t, f i is the fragrance value emitted by the i-th individual, p is the conversion probability, and the optimal solution at time t is g * , Cauchy(0,1) and N(0,1) are the standard Cauchy distribution and the standard Gaussian distribution respectively, r is a random number in [0,1], T2 is the second optimization iteration count, that is, the upper limit of the iteration count of each population in the second optimization, and are the solutions corresponding to the positions of the j-th individual and the k-th individual at time t respectively, and both i and j are natural numbers less than or equal to N.

[0101] The above optimization method is prone to falling into local optimum in the later stage of iteration, which is often caused by some individuals falling into local optimum in a certain dimension during the population update process. Therefore, a vertical crossover operation can be further introduced to improve the ability of the algorithm to jump out of the local optimum. Among them: During the optimization process of the vertical crossover operation, 3 individuals in different dimensions are randomly selected for arithmetic crossover, and the other dimensions remain unchanged, so that the stagnant dimension has the opportunity to jump out of the local optimum. Then the second optimization may further include:

[0102] s21: Randomly select 3 parent individuals in different dimensions.

[0103] s22. Perform a vertical crossover operation on the d1-th, d2-th, and d3-th dimensions of the parental individuals to generate offspring individuals;

[0104] Specifically, the offspring individuals can be obtained through the following formula:

[0105] MS VC (i, d1) = n(qx(i, d1) + (1 - q)x(i, d2)) + (1 - n)x(i, d3)

[0106] where: MS VC (i, d1) is the offspring individual obtained after the vertical operation, q ∈ [0, 1] and n ∈ [0, 1] are random numbers,

[0107] s23. Compare the fitness values of the generated offspring individuals with those of the parental individuals, and retain the individuals with better fitness values as the individuals with the optimal fitness.

[0108] S305. When the current iteration number meets the requirements of the second optimization iteration number and the total iteration number meets the requirements of the initial optimization iteration number, output the optimal individual as the parameters of each layer of the hierarchical prediction large model;

[0109] For example Figure 3 , this step may include:

[0110] s31. Determine whether the current iteration number ≤ the second optimization iteration number T2. If it is less than or equal to, return to step S304 to continue execution. If it is greater than, execute step s32;

[0111] s32. Determine whether the current total iteration number ≤ T0. If it is less than or equal to, randomly select K individuals from the historical optimal individuals in the second optimization of step S304 to replace K individuals in each population in the first optimization of step S303, and return to step S302 to continue optimization. If it is greater than, output the optimal solution.

[0112] Furthermore, after obtaining the optimal solution in each iteration, that is, after executing step s304, the optimal solution can be used as the parameters of the corresponding layer of the hierarchical prediction large model, and the original training data in step S1 can be input into the hierarchical prediction large model for training to verify the effect of the current parameter optimization. Then, execute step S305 for the next iteration optimization, or output the optimal solution to obtain the trained and optimized hierarchical prediction large model.

[0113] S4. Real-time collect power load data and meteorological data and input them into the trained hierarchical prediction large model to predict short-term power load.

[0114] In order to Figure 2Taking the hierarchical prediction large model structure in [specific context] as an example, the real-time collected load data and meteorological data can respectively pass through the NGBoost model, MQSAPN model, RGF model, SLOTH model, iTransformer model, and KAN neural network model connected in parallel in the first layer to output six first prediction results obtained by predicting the load data and meteorological data from different dimensions. Subsequently, the six first prediction results, as well as the collected load data and meteorological data, are respectively subjected to the first fusion processing through the RF model and the lightGBM model connected in parallel in the second layer to obtain two second prediction results. Finally, the six first prediction results output by the first layer and the two second prediction results output by the second layer are input into the PLSR model in the third layer to obtain the final prediction result.

[0115] To prove the prediction accuracy of the short-term power load prediction method based on multi-model fusion provided by the present invention, the power load data and meteorological data in Xi'an area from December 31, 2014 to December 31, 2022 are used for model training and testing. Among them: the data from December 31, 2014 to December 31, 2021 are used for training, and the data after January 1, 2022 are used for testing. Table 1 shows the standard mean absolute error and standard root mean square error of the output results of the present invention and other prediction models.

[0116]

[0117]

[0118] Table 1: Errors of the output results of the present invention and other models

[0119] It can be clearly seen from Table 1 that the short-term power load prediction method based on multi-model fusion of the present invention has smaller errors and higher accuracy than other model prediction methods.

[0120] Furthermore, after obtaining the short-term power load result, a power generation plan can also be formulated according to the short-term power load. The power generation plan can include: grid operation mode and unit maintenance plan. Thereby optimizing resource allocation, reducing operating costs, saving coal and oil, reducing power generation costs, and formulating a reasonable power source construction plan. Therefore, after step S4, the method can further include:

[0121] S5. Determine the power generation amount according to the short-term power load, and configure the grid operation mode and unit maintenance plan based on the power generation amount.

[0122] Based on the above short-term power load prediction method based on multi-model fusion, an embodiment of the present invention also provides a short-term power load prediction device based on multi-model fusion, as Figure 4 described, the device includes:

[0123] The acquisition module 41 is used to respectively acquire power load data and meteorological data to obtain original training data;

[0124] The construction module 42 is used to construct a hierarchical prediction large model. The hierarchical prediction large model includes a first layer, a second layer, and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; at least one sub-model is included in each layer;

[0125] The training and optimization module 43 is used to train and optimize the parameters of each layer in the hierarchical prediction large model based on the original training data;

[0126] The prediction module 44 is used to collect power load data and meteorological data in real time and input them into the trained hierarchical prediction large model to predict short-term power load.

[0127] In one implementation, the first layer includes a plurality of first sub-models connected in parallel, the second layer includes a plurality of second sub-models connected in parallel, and the third layer includes a third sub-model.

[0128] Furthermore, the types of the first sub-models are different, including: the KAN neural network model, and at least one model selected from the natural gradient boosting machine NGBoost, the multi-head quantum self-attention neural network prediction model MQSAPN, the regularized greedy forest RGF, the SLOTH model, and the iTransformer model; the types of the second sub-models are different, and at least two selected from the random forest model RF and the light gradient boosting machine lightGBM; the third sub-model is the partial least squares regression model PLSR.

[0129] The training and optimization module 43 includes:

[0130] The initialization unit is used to initialize the parameter value range of each layer of the hierarchical prediction large model and configure the initial optimization iteration times;

[0131] The configuration unit is used to randomly generate N populations, initialize the fitness of all individuals in the N populations, and configure the first optimization iteration times;

[0132] The first optimization unit is used to respectively perform the first optimization on the N populations based on the first optimization iteration times, update the fitness, and obtain the optimal individuals of the N populations;

[0133] The second optimization unit is used to use the optimal individuals of the N populations as the initial population, configure the second optimization iteration times, and perform the second optimization to obtain the individual with the optimal fitness.

[0134] An output unit, configured to output the optimal individual as the parameters of each layer of the hierarchical prediction large model after the current iteration number meets the requirements of the second optimization iteration number and the total iteration number meets the requirements of the initial optimization iteration number;

[0135] Where: N is a natural number.

[0136] In a specific embodiment, each population contains m individuals. The first optimization unit is configured to perform the following steps for updating and optimizing the m individuals in each population: successively perform triangle topology unit formation, global aggregation, and local aggregation on the m individuals to update the m individuals; calculate the fitness of the m individuals after the update, and update the optimal individual of the current iteration according to the magnitude of the fitness; determine whether the current iteration number reaches the first optimization iteration number; if not satisfied, return to execute the steps of updating and optimizing the m individuals in each population; if satisfied, compare the fitness of the optimal individuals obtained by each population in each layer of iteration, and use the individual with the highest fitness as the optimal individual of the population;

[0137] Where: m is a natural number.

[0138] Optionally, the second optimization unit can be performed by the following formula:

[0139]

[0140] Where: is the position of the i-th individual at time t, f i is the fragrance value emitted by the i-th individual, p is the conversion probability, and the optimal solution at time t is g * , Cauchy(0,1) and N(0,1) are the standard Cauchy distribution and the standard Gaussian distribution respectively, r is a random number in [0,1], T2 is the second optimization iteration number, and are the solutions corresponding to the positions of the j-th individual and the k-th individual at time t respectively, and both i and j are natural numbers less than or equal to N.

[0141] Furthermore, the second optimization unit is further configured to: randomly select 3 parent individuals with different dimensions; perform a vertical crossover operation on the d1-th, d2-th, and d3-th dimensions of the parent individuals to generate offspring individuals; compare the fitness values of the generated offspring individuals with those of the parent individuals, and retain the individuals with better fitness values as the individuals with the optimal fitness.

[0142] Optionally, the offspring individuals are obtained by the following formula:

[0143] MS VC (i, d1) = n(qx(i, d1)+(1 - q)x(i, d2))+(1 - n)x(i, d3)

[0144] Wherein: MS VC (i, d1) is the offspring individual obtained after vertical operation, and q ∈ [0, 1] and n ∈ [0, 1] are random numbers.

[0145] Those skilled in the art can understand that each module in the above device embodiments can be distributed in the device according to the description, or can be correspondingly changed and distributed in one or more devices different from the above embodiments. The modules of the above embodiments can be combined into one module, or can be further split into multiple sub-modules.

[0146] The following describes an embodiment of the electronic device of the present invention. This electronic device can be regarded as an implementation form in physical form of the above method and device embodiments of the present invention. For the details described in the embodiment of the electronic device of the present invention, they should be regarded as a supplement to the above method or device embodiments; for the details not disclosed in the embodiment of the electronic device of the present invention, they can be implemented with reference to the above method or device embodiments.

[0147] Figure 5 is a structural block diagram of an exemplary embodiment of an electronic device according to the present invention. Figure 5 The electronic device shown is only an example and should not bring any limitation to the functions and usage scope of the embodiments of the present invention.

[0148] As Figure 5 shown, the electronic device 400 of this exemplary embodiment is presented in the form of a general-purpose data processing device. The components of the electronic device 400 may include but are not limited to: at least one processing unit 410, at least one storage unit 420, a bus 430 connecting different electronic device components (including the storage unit 420 and the processing unit 410), a display unit 440, etc.

[0149] Wherein, the storage unit 420 stores a computer-readable program, which can be the source program or the code of the read-only program. The program can be executed by the processing unit 410, so that the processing unit 410 executes the steps of various embodiments of the present invention. For example, the processing unit 410 can execute as Figure 1 shown in the steps.

[0150] The bus 430 can represent one or more of several types of bus structures, including a storage unit bus or a storage unit controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any bus structure in a variety of bus structures.

[0151] The electronic device 400 can also communicate with one or more external devices 100 (such as a keyboard, a display, a network device, a Bluetooth device, etc.), enabling a user to interact with the electronic device 400 via these external devices 100, and / or enabling the electronic device 400 to communicate with one or more other data processing devices (such as a router, a modem, etc.). Such communication can be carried out through an input / output (I / O) interface 450, and can also be through a network adapter 460 with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network). The network adapter 460 can communicate with other modules of the electronic device 400 through a bus 430.

[0152] Figure 6 is a schematic diagram of an embodiment of a computer-readable medium of the present invention. As Figure 6 shown, the computer program can be stored on one or more computer-readable media. The computer-readable medium can be a readable signal medium or a readable storage medium. When the computer program is executed by one or more data processing devices, the computer-readable medium can implement the above method of the present invention, that is: respectively collect power load data and meteorological data to obtain original training data; construct a hierarchical prediction large model, the hierarchical prediction large model includes a first layer, a second layer and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; at least one sub-model is included in each layer; optimize the parameters of each layer in the hierarchical prediction large model; use the original training data to train the optimized hierarchical prediction large model; collect power load data and meteorological data in real time and input them into the trained hierarchical prediction large model to predict short-term power load.

[0153] In the above specific embodiments, the purpose, technical solution and beneficial effects of the present invention are further described in detail. It should be understood that the present invention is not inherently related to any specific computer, virtual device or electronic device, and various general-purpose devices can also implement the present invention. The above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A short-term electric load forecasting method based on multi-model fusion, characterized in that Including: Collecting power load data and meteorological data respectively to obtain original training data; The meteorological data includes: temperature, humidity and air pressure; Constructing a hierarchical prediction large model, which includes a first layer, a second layer and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; where: the first layer includes: an NGBoost model, an MQSAPN model, an RGF model, a SLOTH model, an iTransformer model and a KAN neural network model connected in parallel, the second layer includes: an RF model and a lightGBM model connected in parallel, and the third layer only includes one PLSR model; Training and optimizing the parameters of each layer in the hierarchical prediction large model based on the original training data; Collecting power load data and meteorological data in real time and inputting them into the trained hierarchical prediction large model to predict short-term power load.

2. The short-term electric load forecasting method based on multi-model fusion according to claim 1, wherein The training and optimizing the parameters of each layer in the hierarchical prediction large model based on the original training data includes: Initializing the parameter value ranges of each layer of the hierarchical prediction large model and configuring the initial optimization iteration times; Randomly generating N populations, initializing the fitness of all individuals in the N populations, and configuring the first optimization iteration times; Performing the first optimization on the N populations respectively based on the first optimization iteration times, and updating the fitness to obtain the optimal individuals of the N populations; Taking the optimal individuals of the N populations as the initial population, and configuring the second optimization iteration times to perform the second optimization to obtain the individual with the optimal fitness; When the current iteration times meet the requirements of the second optimization iteration times and the total iteration times meet the requirements of the initial optimization iteration times, outputting the optimal individual as the parameters of each layer of the hierarchical prediction large model; Where: N is a natural number.

3. The short-term electric load forecasting method based on multi-model fusion according to claim 2, characterized in that, Each population contains m individuals, and the performing the first optimization on the N populations respectively based on the first optimization iteration times, and updating the fitness to obtain the optimal individuals of the N populations includes: Steps for updating and optimizing the m individuals in each population: successively performing triangular topological unit formation, global aggregation, and local aggregation on the m individuals to update the m individuals; calculating the fitness of the m individuals after updating, and updating the optimal individual of this iteration according to the size of the fitness; Judging whether the current iteration times reach the first optimization iteration times; If not satisfied, returning to execute the steps of updating and optimizing the m individuals in each population; If satisfied, comparing the fitness of the optimal individuals obtained by each population in each layer iteration, and taking the individual with the highest fitness as the optimal individual of the population; Where: m is a natural number.

4. The short-term electric load forecasting method based on multi-model fusion according to claim 2, wherein, The second optimization is carried out through the following formula: Wherein: is the position of the \(i\)-th individual at time \(t\), \(f_i\) is the fragrance value emitted by the \(i\)-th individual, \(p\) is the transition probability, the optimal solution at time \(t\) is \(g^*\), Cauchy(0,1) and N(0,1) are the standard Cauchy distribution and the standard Gaussian distribution respectively, \(r\) is a random number in \([0,1]\), \(T_2\) is the number of times of the second optimization iteration, and are the solutions corresponding to the positions of the \(j\)-th individual and the \(k\)-th individual at time \(t\) respectively, and both \(i\) and \(j\) are natural numbers less than or equal to \(N\).

5. The short-term electric load forecasting method based on multi-model fusion according to claim 4, characterized in that The second optimization further includes: Randomly selecting 3 parent individuals with different dimensions; Performing a vertical crossover operation on the d1-th, d2-th, and d3-th dimensions of the parent individuals to generate offspring individuals; Comparing the fitness values of the generated offspring individuals with those of the parent individuals, and retaining the individuals with better fitness values as the individuals with the optimal fitness.

6. The short-term electric load forecasting method based on multi-model fusion according to claim 5, characterized in that The offspring individuals are obtained through the following formula: MSVC(i, d1) = n(qx(i, d1) + (1 - q)x(i, d2)) + (1 - n)x(i, d3) Wherein: MSVC(i, d1) is the offspring individual obtained after vertical operation, and q ∈ [0, 1] and n ∈ [0, 1] are random numbers.

7. A short-term electric load forecasting device based on multi-model fusion, characterized in that, Including: A collection module for respectively collecting power load data and meteorological data to obtain original training data; The meteorological data includes: temperature, humidity, and air pressure; A construction module for constructing a hierarchical prediction large model, the hierarchical prediction large model includes a first layer, a second layer, and a third layer connected in sequence, and the input end of the first layer is connected to the input end of the second layer, and the output end of the first layer is connected to the input end of the third layer; wherein: the first layer includes: an NGBoost model, an MQSAPN model, an RGF model, a SLOTH model, an iTransformer model, and a KAN neural network model connected in parallel, the second layer includes: an RF model and a lightGBM model connected in parallel, and the third layer only includes one PLSR model; An optimization training module for training and optimizing the parameters of each layer in the hierarchical prediction large model based on the original training data; A prediction module for real-time collecting power load data and meteorological data and inputting them into the trained hierarchical prediction large model to predict short-term power load.

8. An electronic device, including: A processor; And A memory storing computer-executable instructions, the computer-executable instructions, when executed, cause the processor to execute the method according to any one of claims 1-6.

Images