Multi-objective optimization power load interval prediction method and system

Abstract

The present application relates to the technical field of power grid operation and maintenance, and particularly relates to a multi-target optimization power load interval prediction method and system. A number of high-frequency burst feature data intervals are determined by extracting and clustering burst event features according to power grid operation monitoring records of a target power grid in a historical time range; a power load prediction model is obtained based on the number of high-frequency burst feature data intervals; real-time burst feature data of a current power grid burst event is compared with the number of high-frequency burst feature data intervals respectively to obtain a plurality of burst scenario similarities; a model calling strategy is configured according to the plurality of burst scenario similarities, and the power load prediction model is called to predict power load in a preset time zone; an interval fitting strategy is configured according to the plurality of burst scenario similarities, and J initial load prediction intervals are fitted to output a power load prediction interval of the target power grid in the preset time zone. The present application realizes accurate prediction of a power load prediction interval.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid operation and maintenance, and particularly relates to a multi-objective optimization power load interval prediction method and system. BACKGROUND

[0002] In the operation of a power system, power load prediction is an important basis for ensuring the safe, stable and economic operation of a power grid. Accurate load prediction helps to reasonably arrange power generation plans, optimize power grid dispatching and reduce energy loss. However, power load is affected by many factors, especially various types of emergencies, which can cause large fluctuations in load and bring great challenges to prediction work. Traditional power load prediction methods often fail to fully consider the impact of emergencies or are not adaptable to different emergency scenarios, resulting in a significant decrease in prediction accuracy when emergencies occur, which makes it difficult to meet the needs of actual power grid operation. SUMMARY

[0003] The present application provides a multi-objective optimization power load interval prediction method and system to solve the technical problem of insufficient prediction accuracy and difficulty in meeting the needs of actual power grid operation in the prior art.

[0004] The technical solution of the present application to solve the above technical problem is as follows:

[0005] In a first aspect, the present application provides a multi-objective optimization power load interval prediction method, comprising: extracting and clustering emergency features according to power grid operation monitoring records of a target power grid in a historical time range to determine a plurality of high-frequency emergency feature data intervals; constructing a plurality of power load prediction plug-ins based on the plurality of high-frequency emergency feature data intervals, and combining to obtain a power load prediction model; comparing real-time emergency feature data of a current power grid emergency with the plurality of high-frequency emergency feature data intervals respectively to obtain a plurality of emergency scenario similarities; configuring a model calling strategy according to the plurality of emergency scenario similarities, calling the power load prediction model to predict power load in a preset time zone, and outputting J initial load prediction intervals, wherein J is an integer greater than or equal to 1; configuring an interval fusion strategy according to the plurality of emergency scenario similarities, fitting the J initial load prediction intervals, and outputting a power load prediction interval of the target power grid in the preset time zone.

[0006] Optionally, according to the power grid operation monitoring record of the target power grid in the historical time range, the burst event feature extraction and feature clustering are performed to determine a plurality of high-frequency burst feature data intervals, including: according to the power grid operation monitoring record of the target power grid in the historical time range, the burst event feature extraction is performed to obtain a sample burst feature data set, wherein the burst event feature at least includes a weather type, weather data, a power load fluctuation amplitude, a device fault frequency, a voltage fluctuation amplitude, a power system frequency fluctuation amplitude, a power grid load growth rate, an instantaneous load change rate and a load distribution uniformity coefficient; the K-means clustering algorithm is used to cluster the sample burst feature data set to determine a plurality of clustering centers and a plurality of high-frequency burst feature data intervals.

[0007] Optionally, a plurality of power load prediction plug-ins are constructed based on the plurality of high-frequency burst feature data intervals, including: a first high-frequency burst feature data interval is randomly selected, and a sample power load sequence set is collected according to the power grid operation monitoring record of the target power grid in the historical time range, and a historical power load of different sample power load sequences in a future historical time zone is obtained to obtain a sample predicted power load set, with the first high-frequency burst feature data interval as a conditional constraint; the sample power load sequence set and the sample predicted power load set are used as sample training data, and are equally divided into K parts, and K times are selected with replacement, a first training set is constructed, and K times are iteratively selected to obtain K training sets, wherein K is an integer greater than or equal to 10 and less than or equal to 50; the K training sets are used to train a long short-term memory network to convergence, K first power load prediction branches are obtained, a first power load prediction plug-in is constructed by integration, and a plurality of power load prediction plug-ins corresponding to the plurality of high-frequency burst feature data intervals are obtained by training in turn.

[0008] Optionally, according to the multiple burst scene similarity configuration model calling strategy, including: if the number of the multiple burst scene similarities greater than or equal to a first similarity threshold is not 0, selecting a power load prediction plug-in corresponding to the maximum burst scene similarity as an adaptive prediction model; if the number of the multiple burst scene similarities greater than or equal to the first similarity threshold is 0 and the number of the multiple burst scene similarities greater than or equal to a second similarity threshold is not 0, selecting a plurality of power load prediction plug-ins corresponding to the multiple burst scene similarities greater than or equal to the second similarity threshold as the adaptive prediction model, wherein the second similarity threshold is less than the first similarity threshold; if the number of the multiple burst scene similarities greater than or equal to the second similarity threshold is 0, selecting all power load prediction plug-ins in the power load prediction model as the adaptive prediction model.

[0009] The power load prediction in the preset time zone is performed, and J initial load prediction intervals are output, including: collecting a historical power load sequence before a current time node, performing load fluctuation analysis, and setting a branch call number P according to a load fluctuation coefficient; calling the adaptive prediction model, wherein the adaptive prediction model includes J power load prediction plugins; randomly selecting P power load prediction branches in K power load prediction branches of the J power load prediction plugins for prediction, and outputting J initial load prediction intervals, wherein the initial load prediction intervals are constructed based on P prediction results of the P power load prediction branches.

[0010] The load fluctuation analysis is performed, and the branch call number P is set according to the load fluctuation coefficient, including: performing load fluctuation calculation according to the historical power load sequence, and outputting a load fluctuation coefficient, wherein the load fluctuation coefficient is a ratio of a standard deviation of power load to a mean value in the historical power load sequence; taking the load fluctuation coefficient as a prediction complexity, and taking a ratio of the prediction complexity to a historical maximum prediction complexity in a historical time range as a branch selection adjustment coefficient; taking an integer of a product of the branch selection adjustment coefficient and an initial branch number as P, wherein the initial branch number is 5, and P is an integer greater than or equal to 3 and less than or equal to K.

[0011] If J is 1, the initial load prediction interval is set as a power load prediction interval of a target power grid in the preset time zone; if J is not 1, an interval fitting strategy is configured according to the plurality of burst scenario similarities, and the J initial load prediction intervals are fitted, and a power load prediction interval of the target power grid in the preset time zone is output.

[0012] The interval fitting strategy is configured according to the plurality of burst scenario similarities, the J initial load prediction intervals are fitted, and a power load prediction interval of a target power grid in a preset time zone is output, including: obtaining J burst scenario similarities corresponding to the J initial load prediction intervals in the plurality of burst scenario similarities; configuring J fusion confidence weights according to the J burst scenario similarities, wherein the fusion confidence weight is a ratio of a burst scenario similarity to a sum of J burst scenario similarities; setting a corresponding random sampling number according to the J fusion confidence weights, wherein the random sampling number is a product of the fusion confidence weight and a preset sampling number, and the preset sampling number is 1000; fitting the J initial load prediction intervals based on the corresponding random sampling number, and outputting a power load prediction interval.

[0013] The interval fitting of the J initial load prediction intervals is performed based on the corresponding random sampling quantity, and a power load prediction interval is output, including: based on the corresponding random sampling quantity, random sampling with replacement is performed in the J initial load prediction intervals respectively to obtain a sampled power load data set; and the minimum sampled power load data and the maximum sampled power load data in the sampled power load data set are selected to construct a power load prediction interval and output.

[0014] In a second aspect, the present application provides a multi-objective optimization power load interval prediction system, comprising:

[0015] An emergency feature extraction module is configured to perform emergency feature extraction and feature clustering according to power grid operation monitoring records of a target power grid in a historical time range, and determine a plurality of high-frequency emergency feature data intervals.

[0016] A power load prediction model acquisition module is configured to construct a plurality of power load prediction plug-ins based on the plurality of high-frequency emergency feature data intervals, and combine to obtain a power load prediction model.

[0017] An emergency scenario similarity acquisition module is configured to perform similarity comparison between real-time emergency feature data of a current power grid emergency and the plurality of high-frequency emergency feature data intervals respectively, and obtain a plurality of emergency scenario similarities.

[0018] A power load prediction model calling module is configured to configure a model calling strategy according to the plurality of emergency scenario similarities, call the power load prediction model to perform power load prediction in a preset time zone, and output J initial load prediction intervals, wherein J is an integer greater than or equal to 1.

[0019] A power load prediction interval fitting module is configured to configure an interval fusion strategy according to the plurality of emergency scenario similarities, perform interval fitting on the J initial load prediction intervals, and output a power load prediction interval of the target power grid in the preset time zone.

[0020] By implementing the present application, a plurality of high-frequency emergency feature data intervals can be determined by performing emergency feature extraction and feature clustering according to power grid operation monitoring records of a target power grid in a historical time range, which can comprehensively cover all kinds of possible emergency factors that may affect power load, provide accurate feature basis for subsequent power load prediction, and through clustering, high-frequency emergency scenarios can be determined, irrelevant data interference can be reduced, power load prediction can be focused on common emergency situations, and prediction pertinence can be improved.

[0021] By implementing the present application, a plurality of power load prediction plug-ins can be constructed based on the plurality of high-frequency burst feature data intervals, and a power load prediction model can be obtained by combination, wherein each plug-in is specially adapted to a specific high-frequency burst scenario, improving the prediction accuracy for different scenarios, and the plug-in is constructed in an integrated multiple branch manner, reducing the prediction error of a single model, and enhancing the stability and reliability of the power load prediction model.

[0022] By implementing the present application, the real-time burst feature data of the current power grid burst event can be compared with the plurality of high-frequency burst feature data intervals respectively, a plurality of burst scenario similarities can be obtained, and a basis can be provided for subsequent model calling strategy formulation, so that the called prediction plug-in can be adapted to the current actual situation to the greatest extent, and the applicability of prediction is improved.

[0023] By implementing the present application, a model calling strategy can be configured according to the plurality of burst scenario similarities, the power load prediction model can be called to perform power load prediction in a preset time zone, and J initial load prediction intervals can be output, wherein J is an integer greater than or equal to 1. This flexible model calling strategy can select the most suitable prediction plug-in according to the actual similarity, avoiding the blindness of model selection, and the number of branch calls is set in combination with the load fluctuation coefficient, so that the prediction accuracy is guaranteed while the calculation resource consumption is reasonably controlled.

[0024] By implementing the present application, an interval fusion strategy can be configured according to the plurality of burst scenario similarities, the J initial load prediction intervals can be fitted, and a power load prediction interval of the target power grid in the preset time zone can be output. Through weight distribution and sampling fitting, the information of a plurality of initial intervals is integrated, the limitation of a single interval is reduced, the final prediction interval is more accurate and more reliable.

[0025] In summary, by implementing the present application, accurate and robust prediction of power load can be realized, the influence of various burst events on power load can be effectively coped with, and the accuracy and adaptability of power load prediction are improved. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A flowchart of a multi-objective optimization power load interval prediction method provided by the present application is shown.

[0027] Figure 2 A structural diagram of a multi-objective optimization power load interval prediction system provided by the present application is shown.

[0028] In the drawings, the components represented by the numbers are as follows:

[0029] The emergency event feature extraction module 11, the power load prediction model acquisition module 12, the emergency event scene similarity acquisition module 13, the power load prediction model calling module 14, and the power load prediction interval fitting module 15. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0031] In the description of the present application, the terms "first" and "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" can explicitly or implicitly include one or more features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0032] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or explanation". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail in order to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope of the principles and characteristics disclosed.

[0033] Embodiment one, as shown in the present application, provides a multi-objective optimization power load interval prediction method, comprising: Figure 1

[0034] S100: According to the power grid operation monitoring record of the target power grid in the historical time range, emergency event feature extraction and feature clustering are performed, and a plurality of high-frequency emergency feature data intervals are determined;

[0035] S200: Based on the plurality of high-frequency emergency feature data intervals, a plurality of power load prediction plug-ins are constructed, and a power load prediction model is obtained by combination;

[0036] ​S300: Similarity comparison is performed between the real-time burst feature data of the current power grid burst and the plurality of high-frequency burst feature data intervals, to obtain a plurality of burst scene similarities;

[0037] S400: A model calling strategy is configured according to the plurality of burst scene similarities, the power load prediction model is called to perform power load prediction in a preset time zone, and J initial load prediction intervals are output, where J is an integer greater than or equal to 1;

[0038] S500: An interval fusion strategy is configured according to the plurality of burst scene similarities, interval fitting is performed on the J initial load prediction intervals, and a power load prediction interval of the target power grid in the preset time zone is output.

[0039] In step S100 of the embodiment of the present application, burst event feature extraction and feature clustering are performed according to power grid operation monitoring records of the target power grid in a historical time range, to determine a plurality of high-frequency burst feature data intervals, including:

[0040] Burst event feature extraction is performed according to power grid operation monitoring records of the target power grid in a historical time range, to obtain a sample burst feature data set, where the burst event features at least include weather type, weather data, power load fluctuation amplitude, equipment fault frequency, voltage fluctuation amplitude, power system frequency fluctuation amplitude, power grid load growth rate, instantaneous load change rate, and load distribution uniformity coefficient;

[0041] The K-means clustering algorithm is used to cluster the sample burst feature data set, to determine a plurality of cluster centers and a plurality of high-frequency burst feature data intervals.

[0042] In the embodiment of the present application, burst event feature extraction and feature clustering are performed on power grid operation monitoring records, the purpose of which is to mine different types of high-frequency burst feature data intervals from historical power grid operation data, to provide feature templates for subsequent construction of load prediction models adapted to various scenes and accurate identification of current burst scenes. For example, extreme weather, equipment failure, and power consumption peak, these high-frequency burst situations, the feature intervals of which in temperature, load fluctuation, and other dimensions are determined in advance, so that when a real-time event occurs, the corresponding scene can be quickly matched, and the prediction is more in line with the actual situation.

[0043] First, the characteristics of the incident need to be extracted from the grid operation monitoring records of the target grid in the historical time range, and the sample burst characteristic data set is obtained. That is, from the grid operation monitoring records of the target grid in the historical time range, the weather type, weather data, power load fluctuation amplitude, equipment failure frequency, voltage fluctuation amplitude, power system frequency fluctuation amplitude, grid load growth rate, instantaneous load change rate and load distribution uniformity coefficient and other characteristics related to the incident are extracted to form a data set.

[0044] Among them, the weather type refers to the weather category that can affect the grid load, such as common sunny, rainy, snowy, high temperature, cold wave, typhoon, etc.

[0045] Weather data is a quantitative index that specifically describes the weather conditions, including temperature, humidity, wind speed, wind direction, etc.

[0046] The power load fluctuation amplitude is the difference between the actual load and the reference load in a certain time, such as 1 day, accounting for the percentage of the reference load, reflecting the degree of load change. For example, the reference load of a regional power grid before the morning peak is 500 MW, and the actual load rises to 600 MW at the morning peak, then the load fluctuation amplitude is |(600-500)| / 500x100%=20%.

[0047] The equipment failure frequency refers to the number of times the grid equipment such as transformers, circuit breakers, and transmission lines fails within a certain time, such as 1 month.

[0048] The voltage fluctuation amplitude is the percentage of the actual voltage deviating from the rated value, reflecting the voltage stability, voltage fluctuation amplitude=|(actual voltage-rated voltage)| / rated voltagex100%. Voltage fluctuation will affect the normal operation of electrical equipment and also affect the load characteristics of the power grid.

[0049] The power system frequency fluctuation amplitude is the difference between the actual frequency and the rated frequency of the power system, accounting for the percentage of the rated frequency, where the rated frequency of the power system in China is 50Hz. Then the power system frequency fluctuation amplitude=|(actual frequency-50)| / 50x100%.

[0050] The grid load growth rate is the ratio of load growth to initial load within a certain time interval, reflecting the speed of load growth, grid load growth rate=[(final load-initial load) / initial load]x100%. For example, a new residential area has a low occupancy rate in January, with a grid load of 2MW, and the occupancy rate increases in March, with a grid load of 6MW. The load growth rate within 2 months is [(6-2) / 2]x100%=200%. High grid load growth rate indicates that subsequent electricity demand continues to increase, and the power grid needs to be planned for capacity expansion in advance, which also affects short-term load forecasting and scheduling.

[0051] The instantaneous load change rate is the ratio of the load change amount to the initial load in a very short time, such as 1 minute or a few seconds, which reflects the instantaneous rate of load change and is understood as the degree of severity of load change in a short time. The instantaneous load change rate = [load change amount / (initial load x time interval)] x 100%. For example, when a large motor in a factory is started, the grid load increases from 10 MW to 15 MW in 1 second, and the instantaneous load change rate = [(15-10) / (10x1)]x100% = 50% / second.

[0052] The load distribution uniformity coefficient is used to measure the balance degree of load distribution in different regions and different time periods of the power grid. It is generally calculated by the deviation of the load of each region or each time period from the average load. There are many specific calculation methods, all of which are prior art and will not be described here. It can be simply understood as reflecting whether the load is concentrated in space or time. A high uniformity coefficient indicates that the load distribution is dispersed and balanced, and vice versa. For example, if the load distribution uniformity coefficient of a certain power grid region is 50%, which is lower than the average value of 65% of multiple power grid regions, a low load distribution uniformity coefficient means that the local load of the power grid is too heavy, and needs to be strengthened or increased in capacity, which also affects the accuracy of load forecasting.

[0053] According to the power grid operation monitoring records of the target power grid in the historical time range, the sample burst feature data set is obtained, wherein the historical time range can be the past 1 year, and the burst event features at least include the above-mentioned weather type, weather data, power load fluctuation amplitude, equipment failure frequency, voltage fluctuation amplitude, power system frequency fluctuation amplitude, power grid load growth rate, instantaneous load change rate and load distribution uniformity coefficient.

[0054] Next, a clustering algorithm such as K-means algorithm is used to cluster the sample burst feature data set. The clustering algorithm calculates the similarity between the feature data, classifies similar features into a class, and determines the cluster center and interval. For example, the clustering result of the sample burst feature data set contains three types of event data: extreme weather, equipment failure and peak electricity consumption. The feature data in each cluster is distributed around the cluster center, forming feature intervals in dimensions such as weather data and equipment failure frequency. For example, after clustering the equipment failure event, the temperature data interval is 25-30℃, and the equipment failure frequency is 5-6 times, etc. to determine the high-frequency burst feature data interval of various clusters.

[0055] For example, assuming that the target power grid has a record of power grid operation monitoring in a historical time range, there are multiple occurrences of high temperature of about 32°C, wind speed of 18 m / s, and load fluctuation of about 12% in summer. These data are clustered into extreme weather events because of similar characteristics. After algorithm calculation, it is determined that the characteristic interval of this type of temperature is 30-35°C, and the characteristic interval of wind speed is 15-20 m / s. The characteristic range of this high-frequency burst scenario of extreme weather is clearly defined. When real-time high temperature and strong wind weather occurs subsequently, it can be judged whether it belongs to this scenario by comparing these intervals, thereby assisting load prediction.

[0056] The clustering logic based on temperature, wind speed, and other characteristics described above is only used to illustrate the principle of clustering the sample burst characteristic data set described herein. In the actual clustering process, multiple dimensions of characteristics such as the above meteorological types, meteorological data, power load fluctuation amplitude, equipment fault frequency, voltage fluctuation amplitude, power system frequency fluctuation amplitude, power grid load growth rate, instantaneous load change rate, and load distribution uniformity coefficient need to be considered for accurate clustering.

[0057] In step S200 of the embodiments of the present application, a plurality of power load prediction plug-ins are constructed based on the plurality of high-frequency burst characteristic data intervals, including:

[0058] A first high-frequency burst characteristic data interval is randomly selected. Based on the first high-frequency burst characteristic data interval as a conditional constraint, a sample power load sequence set is collected according to the power grid operation monitoring record of the target power grid in a historical time range, and historical power loads of different sample power load sequences in a future historical time zone are obtained to obtain a sample predicted power load set;

[0059] The sample power load sequence set and the sample predicted power load set are used as sample training data and are equally divided into K parts. K times are selected with replacement. A first training set is constructed. K times are selected iteratively to obtain K training sets, wherein K is an integer greater than or equal to 10 and less than or equal to 50;

[0060] The K training sets are used to train a long short-term memory network to convergence, respectively, to obtain K first power load prediction branches. A first power load prediction plug-in is constructed by integration, and a plurality of power load prediction plug-ins corresponding to the plurality of high-frequency burst characteristic data intervals are obtained by training in turn.

[0061] In the embodiments of the present application, a plurality of power load prediction plug-ins are constructed to provide a special tool for subsequent response to various power grid burst conditions and accurate prediction of power load, thereby improving the coverage and prediction accuracy of the overall prediction model for complex burst scenarios.

[0062] Specifically, for each high-frequency burst feature data interval, such as extreme weather, equipment failure and other scenarios, a dedicated power load prediction plug-in is constructed. The power load prediction model can accurately fit the load change law under this scenario, solving the problem that the traditional single model cannot adapt to various burst scenarios. For example, temperature, wind speed and other factors have a unique impact on load during extreme weather. The dedicated power load prediction plug-in can focus on such influences to make accurate predictions.

[0063] First, a first high-frequency burst feature data interval is randomly selected. Based on the first high-frequency burst feature data interval as a conditional constraint, a sample power load sequence set is collected according to the power grid operation monitoring records of the target power grid in the historical time range, and the historical power load of different sample power load sequences in the future historical time zone is obtained to obtain a sample prediction power load set;

[0064] That is, according to the characteristics of the first high-frequency burst feature data interval, such as temperature 30-35℃ and wind speed 15-20m / s in the extreme weather scenario, the time period data that meets the interval condition is selected from the power grid operation monitoring records of the target power grid in the historical time range, and the corresponding sample power load sequence set is collected, that is, the continuous load data sequence in these time periods. At the same time, the historical power load of these sequences in the future historical time zone, such as the next 1 hour or 24 hours, is obtained to form a sample prediction power load set. For example, in the extreme weather scenario, the dates in the power grid operation monitoring records in the historical time range that meet the first high-frequency burst feature data interval in terms of temperature, wind speed, etc. are selected, and the power load sequence at each time of the day and the historical power load in the future historical time zone, i.e. the sample prediction power load, are extracted as the input and output for the power load prediction plug-in training.

[0065] Further, a diversified training set needs to be constructed, that is, the sample training data including the sample power load sequence set and the sample prediction power load set is divided into K parts, such as K=20. Then, bootstrap sampling or other sampling with replacement is used to construct a training set by randomly selecting K parts of data each time, and the process is repeated K times to obtain K training sets. For example, 20 data, some of which may be repeatedly selected in each sampling to construct 20 different but overlapping training sets.

[0066] The purpose of the above-mentioned construction of a diversified training set is to increase the diversity of data, so that the subsequent multiple long short-term memory networks (LSTM) branches can learn the feature relationships of the data from different angles. Avoiding the problem of incomplete learning caused by the limited data distribution of a single training set.

[0067] The K training sets are used to train the long short-term memory network to convergence, respectively, to obtain K first power load prediction branches, which are integrated to construct a first power load prediction plug-in, and the K training sets are used to train the long short-term memory network to convergence, respectively, to obtain K first power load prediction branches, which are integrated to construct a first power load prediction plug-in.

[0068] Since the long short-term memory network (LSTM) is suitable for processing the time sequence dependence of the power load sequence, the multi-branch integration can cover different feature sub-scenarios within the interval, therefore, in the embodiments of the present application, an LSTM branch integration scheme is adopted to construct the power load prediction plug-in.

[0069] The training sample source of the power load prediction plug-in is the K training sets obtained in the foregoing steps, including the sample power load sequence set and the sample predicted power load set. The input of the power load prediction plug-in is a power load sequence, such as 24-hour power load values, corresponding to 1-day time sequence data; and the output is a predicted power load, such as predicted actual power load values in the next 24 hours.

[0070] The structure of a single LSTM branch in the power load prediction plug-in includes an input layer, a hidden layer and an output layer. The dimension of the input layer = the length of the power load sequence + the dimension of the auxiliary features, such as taking 24-hour power load, the dimension is 24, and the dimension of the auxiliary features is the dimension of the feature data such as weather data and equipment failure frequency actually considered in the range, such as the dimension of the auxiliary features can be 8;

[0071] The hidden layer is a 2-layer LSTM, and the number of units in each layer = 64, to balance the accuracy and the amount of calculation, and adapt to the time sequence data scale of the power grid;

[0072] The dimension of the output layer = the time zone length of the predicted power load, such as 24 for the predicted 24 hours.

[0073] Then, the K training sets are needed to be used to train the long short-term memory network to convergence respectively, to obtain K first power load prediction branches, which are integrated to construct a first power load prediction plug-in.

[0074] The number of training rounds of the K first power load prediction branches is set to a maximum of 50 rounds, to avoid overfitting, and each round is used to train K branches with 1 training set;

[0075] When the mean absolute error (MAE) of the validation set of a certain LSTM branch is continuously reduced by ≤0.1% for 5 rounds, or the loss function (MSE) is stable in the ±1% interval of the historical minimum value, it is determined that the corresponding LSTM branch converges, and K first power load prediction branches are finally trained.

[0076] Further, the K branches are integrated to form a first power load prediction plug-in. For example, the branch outputs are fused in a simple average, weighted average or the like, or the branches are directly called in parallel, and the results of multiple branches are integrated to make a prediction. Through integration, the complementarity of the multiple branches is utilized to improve the prediction accuracy and stability of the plug-in in the corresponding high-frequency scenario.

[0077] For example, if a weighted average fusion is adopted, weights can be assigned according to the prediction error (MAE) of the K branches on the validation set, and the branch with a smaller error has a higher weight. For example, if the MAE of branch 1 is 0.1 and the MAE of branch 2 is 0.2, the weight ratio is 2:1. In this way, the prediction results of the K branches are weighted and summed to obtain the final power load prediction result. For example, if there are two branches A and B, the weight of branch A is 60% and the prediction value is 500MW by the above method, and the weight of branch B is 40% and the prediction value is 520MW, then the final power load prediction result is 500x60%+520x40%=508MW.

[0078] According to the above process of collecting samples, constructing a diversified training set, training LSTM branches, and integrating, for each high-frequency burst feature data interval, a corresponding power load prediction plug-in is constructed. Each plug-in is trained by a dedicated scene sample and integrated by multiple branches to adapt to the load prediction needs of the corresponding scene.

[0079] In step S300 of the embodiment of the present application, the real-time burst feature data of the current power grid burst event is compared with the plurality of high-frequency burst feature data intervals respectively to obtain a plurality of burst scene similarities.

[0080] The purpose of step S300 is to determine the matching degree of the current power grid burst event scene and the historical high-frequency burst scene according to the plurality of high-frequency burst feature data intervals, and to provide a basis for selecting an adaptive power load prediction plug-in for subsequent use.

[0081] First, the real-time feature data of the current power grid burst event needs to be collected. These features are consistent with the sample burst features extracted in S100, and at least include the weather type, weather data, power load fluctuation amplitude, equipment failure frequency, voltage fluctuation amplitude, power system frequency fluctuation amplitude, power grid load growth rate, instantaneous load change rate, and load distribution uniformity coefficient.

[0082] For example, the real-time feature data of the current power grid burst event can be weather type = heavy rain, temperature = 25℃, wind speed = 8m / s, load fluctuation amplitude = 15%, equipment failure frequency = 2 times / hour, etc.

[0083] Next, the similarity between the real-time burst feature data and each high-frequency burst feature data interval needs to be calculated.

[0084] The similarity calculation can be based on the matching degree of the feature dimensions: for example, if the temperature interval of a certain high-frequency burst feature data interval is 20-30℃, the wind speed interval is 5-10 m / s, and the current real-time temperature is 25℃, the wind speed is 8 m / s, then the two dimension data match, and the corresponding feature similarity value is 1; if the real-time load fluctuation amplitude is 15%, and the load fluctuation amplitude interval of the high-frequency burst feature data interval is 5%-10%, then the dimension data does not match, and the corresponding feature similarity value is 0.

[0085] Finally, a plurality of feature similarity values are obtained, and after all are added and divided by the number of all feature categories participating in comparison, the burst scene similarity can be obtained, for example, the value of the plurality of feature similarity values after addition is 7, and the number of all feature categories participating in comparison is 10, then the burst scene similarity = 7 / 10 = 0.7, that is, 70%.

[0086] Through the above method, the real-time burst feature data of the current power grid burst event is compared with the plurality of high-frequency burst feature data intervals respectively, and a plurality of burst scene similarities are obtained.

[0087] In step S400 of the embodiment of the present application, a model calling strategy is configured according to the plurality of burst scene similarities, including:

[0088] If the number of the plurality of burst scene similarities greater than or equal to the first similarity threshold is not 0, then the power load prediction plug-in corresponding to the maximum burst scene similarity is selected as the adaptive prediction model.

[0089] If the number of the plurality of burst scene similarities greater than or equal to the first similarity threshold is 0 and the number of the plurality of burst scene similarities greater than or equal to the second similarity threshold is not 0, then the plurality of power load prediction plug-ins corresponding to the plurality of burst scene similarities greater than or equal to the second similarity threshold are selected as the adaptive prediction model, wherein the second similarity threshold is less than the first similarity threshold.

[0090] If the number of the plurality of burst scene similarities greater than or equal to the second similarity threshold is 0, then all the power load prediction plug-ins in the power load prediction model are selected as the adaptive prediction model.

[0091] In the embodiment of the present application, the power load prediction plug-in calling strategy is configured according to the plurality of burst scene similarities, the purpose of which is to flexibly select the most suitable power load prediction plug-in according to the similarity of the current burst scene and the historical high-frequency burst scene, to ensure that the called power load prediction plug-in can adapt to the current actual situation to the greatest extent, and to provide a reliable model basis for subsequent accurate output of the initial load prediction interval.

[0092] First, the first similarity threshold and the second similarity threshold need to be preset according to the actual operation demand of the power grid.

[0093] The first similarity threshold is a higher threshold, such as 80%, representing that the current burst scenario is highly matched with a certain historical high-frequency burst scenario;

[0094] The second similarity threshold is a lower threshold, such as 50%, representing that the current burst scenario is not highly matched with part of the historical high-frequency burst scenario, but there is a certain correlation, wherein the second similarity threshold < the first similarity threshold.

[0095] By setting different first and second similarity thresholds, the adaptive prediction plug-ins can be screened in different cases, and the prediction deviation caused by a single model selection method can be avoided.

[0096] Then, the power load prediction plug-in calling strategy needs to be configured in different cases.

[0097] The first case is that one or more burst scenario similarities are greater than or equal to the first similarity threshold, such as 80%, and the power load prediction plug-in corresponding to the maximum burst scenario similarity is selected as the adaptive prediction model.

[0098] The second case is that all burst scenario similarities are less than the first similarity threshold, such as 80%, but one or more burst scenario similarities are greater than or equal to the second similarity threshold, such as 50%, and the power load prediction plug-ins corresponding to these burst scenario similarities are selected as the adaptive prediction model.

[0099] The third case is that all burst scenario similarities are less than the second similarity threshold, such as 50%, indicating that the current scenario has a low correlation with the known historical high-frequency burst scenario, and all power load prediction plug-ins in the power load prediction model are selected as the adaptive prediction model.

[0100] Through the above hierarchical strategy, the power load prediction plug-in dynamic selection based on the burst scenario similarity is realized in this step S400, which not only ensures the prediction accuracy in the high-matching scenario, but also takes into account the prediction comprehensiveness in the low-matching scenario, and provides strong model support for the subsequent generation of the initial load prediction interval.

[0101] In the step S400 of the embodiments of the present application, the power load prediction in the preset time zone is performed, and J initial load prediction intervals are output, including:

[0102] The historical power load sequence before the current time node is collected, the load volatility analysis is performed, and the number P of branch calls is set according to the load fluctuation coefficient;

[0103] The adaptive prediction model is called, wherein the adaptive prediction model includes J power load prediction plug-ins;

[0104] Randomly select P power load prediction branches in the K power load prediction branches of the J power load prediction plugins for prediction, and output J initial load prediction intervals, wherein the initial load prediction intervals are constructed based on P prediction results of the P power load prediction branches.

[0105] In step S400 of the embodiment of the present application, load fluctuation analysis is performed, and the number P of branch calls is set according to the load fluctuation coefficient, including:

[0106] Load fluctuation calculation is performed according to the historical power load sequence, and a load fluctuation coefficient is output, wherein the load fluctuation coefficient is the ratio of the standard deviation of power load to the mean value in the historical power load sequence.

[0107] The load fluctuation coefficient is taken as a prediction complexity, and the ratio of the prediction complexity to the historical maximum prediction complexity in the historical time range is taken as a branch selection adjustment coefficient.

[0108] The product of the branch selection adjustment coefficient and the initial branch number is rounded to P, wherein the initial branch number is 5, and P is an integer greater than or equal to 3 and less than or equal to K.

[0109] In step S400 of the embodiment of the present application, the number P of branch calls is set according to the load fluctuation coefficient, which aims to dynamically adjust the number of LSTM branch calls in the power load prediction plugin according to the fluctuation intensity of the current power load, so as to reasonably control the consumption of computing resources while ensuring the prediction accuracy.

[0110] First, the load fluctuation coefficient needs to be calculated according to the historical power load sequence. The load fluctuation coefficient is specifically calculated as follows: load fluctuation coefficient = standard deviation of power load in the historical power load sequence ÷ mean value of the historical power load sequence.

[0111] In the above formula, the standard deviation of power load in the historical power load sequence reflects the dispersion degree of the historical power load data, that is, the fluctuation amplitude, and the mean value of the historical power load sequence reflects the overall level of the power load. The ratio of the two can eliminate the power load magnitude influence and objectively reflect the relative intensity of power load fluctuation.

[0112] For example, a certain historical power load sequence is [100, 120, 90, 130, 80] (unit: MW), the mean of the power load values in the sequence is 104, the standard deviation is approximately 21.54, and the corresponding load fluctuation coefficient is 21.54 ÷ 104 ≈ 0.21, so the load fluctuation is small. For another sequence [50, 150, 30, 170, 20], the mean of the power load values in the sequence is 84, the standard deviation is approximately 68.77, and the corresponding load fluctuation coefficient is 68.77 ÷ 84 ≈ 0.82, so the load fluctuation is large.

[0113] Further, the load fluctuation coefficient is directly taken as the prediction complexity, and the ratio of the prediction complexity to the historical maximum prediction complexity in the historical time range is taken as the branch selection adjustment coefficient. The historical maximum prediction complexity refers to the maximum load fluctuation coefficient calculated in all power load sequences in the historical time range, such as 1 year.

[0114] The branch selection adjustment coefficient is calculated as follows: branch selection adjustment coefficient = current prediction complexity ÷ historical maximum prediction complexity. For example, assuming that the current load fluctuation coefficient is 0.82 and the historical maximum load fluctuation coefficient is 1.0, the branch selection adjustment coefficient is 0.82 ÷ 1.0 = 0.82; if the current maximum load fluctuation coefficient is 0.21, the branch selection adjustment coefficient is 0.21 ÷ 1.0 = 0.21.

[0115] Further, the product of the branch selection adjustment coefficient and the initial branch number is rounded to P, where the initial branch number is 5 and P is an integer greater than or equal to 3 and less than or equal to K. That is, P = (branch selection adjustment coefficient × initial branch number), the result is rounded to an integer, and it needs to satisfy 3 ≤ P ≤ K, where K is the total number of branches trained in S200, such as 20.

[0116] Through the above steps, the number of branch calls is dynamically adjusted according to the load fluctuation, which reduces the calculation amount while meeting the basic accuracy.

[0117] Further, the adaptive prediction model needs to be called, that is, according to the power load prediction plug-in calling strategy determined in the foregoing steps, the adaptive prediction model corresponding to the current scenario is determined, and the adaptive prediction model includes J power load prediction plug-ins.

[0118] Then, for each of the J power load prediction plug-ins, P power load prediction branches are randomly selected from the K power load prediction branches it contains, such as K = 20. Where P has been calculated by the foregoing steps, and its value range is 3 ≤ P ≤ K.

[0119] For example, if P=4 and K=20, each power load forecasting plug-in randomly selects 4 power load forecasting branches from the 20 power load forecasting branches to perform prediction, and outputs J initial load forecasting intervals.

[0120] Each selected power load forecasting branch outputs a power load forecasting result for a preset time zone, such as the next 24 hours. The prediction results of the P power load forecasting branches of a single power load forecasting plug-in are used to construct an initial load forecasting interval corresponding to the power load forecasting plug-in by a statistical method. For example, the minimum value of the P results is taken as the lower limit, and the maximum value is taken as the upper limit.

[0121] For example, the 4 power load forecasting branches of a certain power load forecasting plug-in predict the power load in the next 24 hours to be 500 MW, 510 MW, 495 MW, and 505 MW, respectively. The initial load forecasting interval of the power load forecasting plug-in is [495 MW, 510 MW].

[0122] The above operation is repeated, and J power load forecasting plug-ins output J initial load forecasting intervals, which provide original data for subsequent initial load forecasting interval fusion.

[0123] In step S400 of the embodiment of the present application, if J is 1, the initial load forecasting interval is set as the power load forecasting interval of the target power grid in the preset time zone. Since J=1, no multi-interval fusion is needed, and the initial load forecasting interval constructed above is directly output as the final power load forecasting interval of the target power grid in the preset time zone.

[0124] If J is not 1, an interval fusion strategy is configured according to the plurality of burst scenario similarities, the J initial load forecasting intervals are fitted, and a power load forecasting interval of the target power grid in the preset time zone is output.

[0125] In step S500 of the embodiment of the present application, when J is not 1, an interval fusion strategy needs to be configured according to the plurality of burst scenario similarities, the J initial load forecasting intervals are fitted, and a power load forecasting interval of the target power grid in the preset time zone is output, including:

[0126] Obtaining J burst scenario similarities corresponding to the J initial load forecasting intervals in the plurality of burst scenario similarities;

[0127] Configuring J fusion confidence weights according to the J burst scenario similarities, wherein the fusion confidence weight is the ratio of a burst scenario similarity to the sum of the J burst scenario similarities;

[0128] According to the J fusion confidence weights, corresponding random sampling quantities are set, wherein the random sampling quantity is the product of the fusion confidence weight and a preset sampling quantity, and the preset sampling quantity is 1000;

[0129] Based on the corresponding random sampling quantities, interval fitting is performed on the J initial load prediction intervals, and a power load prediction interval is output.

[0130] In the embodiments of the present application, the purpose of interval fitting on the J initial load prediction intervals is to generate a comprehensive and reliable final prediction interval by fusing multiple initial load prediction intervals, thereby solving the problem of integrating the prediction results of multiple power load prediction plugins.

[0131] Firstly, the J burst scenario similarities corresponding to the J initial load prediction intervals in the multiple burst scenario similarities need to be obtained. Specifically, J burst scenario similarity values corresponding to the J initial load prediction intervals can be extracted from the multiple burst scenario similarities obtained from S300. For example, when J=2, the corresponding power load prediction plugin 1 and power load prediction plugin 2 are obtained. The burst scenario similarity corresponding to the power load prediction plugin 1 is 80%, and the burst scenario similarity corresponding to the power load prediction plugin 2 is 60%.

[0132] Then, J fusion confidence weights need to be configured according to the J burst scenario similarities, wherein the fusion confidence weight is the ratio of the burst scenario similarity to the sum of the J burst scenario similarities. For example, when J=2, the burst scenario similarities of the corresponding power load prediction plugins are 80% and 60% respectively, and the sum of the burst scenario similarities is 140%. Then the fusion confidence weight of the power load prediction plugin 1 is 80% ÷ 140% ≈ 57.1%, and the fusion confidence weight of the power load prediction plugin 2 is 60% ÷ 140% ≈ 42.9%.

[0133] Further, the initial load prediction intervals of the power load prediction plugins corresponding to the J fusion confidence weights need to be set with corresponding random sampling quantities one by one, wherein each fusion confidence weight corresponds to a random sampling quantity, and the random sampling quantity is the product of the fusion confidence weight and a preset sampling quantity, and the preset sampling quantity is 1000.

[0134] Wherein, the random sampling quantity of the initial load prediction interval corresponding to each power load prediction plugin = the fusion confidence weight of the power load prediction plugin × 1000. In the above example, the random sampling quantity of the initial load prediction interval corresponding to the power load prediction plugin 1 = 57.1% × 1000 ≈ 571, and the random sampling quantity of the initial load prediction interval corresponding to the power load prediction plugin 2 = 42.9% × 1000 ≈ 429, and the sum of the two is 1000.

[0135] In step S500 of this application embodiment, it is also necessary to perform interval fitting on the J initial load prediction intervals based on the corresponding random sampling quantity, and output the power load prediction interval, including:

[0136] Based on the corresponding number of random samples, random sampling with replacement is performed in each of the J initial load prediction intervals to obtain a sampled power load dataset.

[0137] The minimum and maximum sampled power load data in the sampled power load dataset are selected to construct and output the power load prediction interval.

[0138] Specifically, based on the aforementioned sampling quantity, random sampling with replacement needs to be performed on each initial load prediction interval. For example, 571 values ​​are randomly selected from the initial interval [490, 510] of power load prediction plugin 1, and 429 values ​​are randomly selected from the initial interval [480, 500] of power load prediction plugin 2, forming a sampled power load dataset containing 1000 values.

[0139] Then, the minimum value is selected from the sampled power load dataset as the lower limit of the power load prediction interval, and the maximum value is selected as the upper limit of the power load prediction interval. For example, if the minimum value in 1000 sampled power load data is 482 and the maximum value is 508, then the final power load prediction interval is [482, 508].

[0140] Through the above process, the credibility of different power load forecasting plugins is reflected by integrating credibility weights, and information from multiple initial load forecasting intervals is integrated by large-sample random sampling. The final output power load forecasting interval takes into account both bias and comprehensiveness, achieving accurate and efficient forecasting of power load.

[0141] Example 2, as Figure 2 As shown, based on the same inventive concept as the multi-objective optimized power load interval prediction method provided in Embodiment 1, this embodiment of the invention also provides a multi-objective optimized power load interval prediction system, including:

[0142] The sudden event feature extraction module 11 is used to extract and cluster sudden event features based on the power grid operation monitoring records of the target power grid within a historical time range, and to determine several high-frequency sudden event feature data intervals.

[0143] The power load prediction model acquisition module 12 is used to construct several power load prediction plug-ins based on the several high-frequency burst characteristic data intervals, and combine them to obtain a power load prediction model.

[0144] The burst scene similarity obtaining module 13 is configured to compare the real-time burst feature data of the current power grid burst event with the plurality of high-frequency burst feature data intervals respectively to obtain a plurality of burst scene similarities.

[0145] The power load prediction model calling module 14 is configured to configure a model calling strategy according to the plurality of burst scene similarities, call the power load prediction model to perform power load prediction in a preset time zone, and output J initial load prediction intervals, where J is an integer greater than or equal to 1.

[0146] The power load prediction interval fitting module 15 is configured to configure an interval fusion strategy according to the plurality of burst scene similarities, fit the J initial load prediction intervals, and output a power load prediction interval of the target power grid in the preset time zone.

[0147] Further, the burst event feature extraction module 11 includes the following execution steps:

[0148] The burst event feature extraction is performed according to the power grid operation monitoring records of the target power grid in a historical time range to obtain a sample burst feature data set, where the burst event features at least include a weather type, weather data, a power load fluctuation amplitude, a device fault frequency, a voltage fluctuation amplitude, a power system frequency fluctuation amplitude, a power grid load growth rate, an instantaneous load change rate, and a load distribution uniformity coefficient.

[0149] The K-means clustering algorithm is used to cluster the sample burst feature data set to determine a plurality of clustering centers and a plurality of high-frequency burst feature data intervals.

[0150] Further, the power load prediction model obtaining module 12 includes the following execution steps:

[0151] A first high-frequency burst feature data interval is randomly selected, and a sample power load sequence set is collected and historical power loads of different sample power load sequences in a future historical time zone are obtained according to the power grid operation monitoring records of the target power grid in the historical time range, with the first high-frequency burst feature data interval as a conditional constraint, to obtain a sample predicted power load set.

[0152] The sample power load sequence set and the sample predicted power load set are used as sample training data and are equally divided into K parts, K times are selected with replacement, a first training set is constructed, K times are iteratively selected, and K training sets are obtained, where K is an integer greater than or equal to 10 and less than or equal to 50.

[0153] The K training sets are used to train long short-term memory networks to convergence respectively, K first power load prediction branches are obtained, a first power load prediction plug-in is constructed by integration, and a plurality of power load prediction plug-ins corresponding to a plurality of high-frequency burst feature data intervals are obtained by training in turn.

[0154] Further, the power load prediction model calling module 14 comprises the following execution steps:

[0155] If the number of the plurality of burst scenario similarities greater than or equal to the first similarity threshold is not 0, the power load prediction plug-in corresponding to the maximum burst scenario similarity is selected as the adaptive prediction model;

[0156] If the number of the plurality of burst scenario similarities greater than or equal to the first similarity threshold is 0 and the number of the plurality of burst scenario similarities greater than or equal to the second similarity threshold is not 0, the plurality of power load prediction plug-ins corresponding to the plurality of burst scenario similarities greater than or equal to the second similarity threshold are selected as the adaptive prediction model, wherein the second similarity threshold is less than the first similarity threshold;

[0157] If the number of the plurality of burst scenario similarities greater than or equal to the second similarity threshold is 0, all the power load prediction plug-ins in the power load prediction model are selected as the adaptive prediction model.

[0158] A historical power load sequence before a current time node is collected, load fluctuation analysis is performed, and the number P of branch calls is set according to a load fluctuation coefficient;

[0159] The adaptive prediction model is called, wherein the adaptive prediction model comprises J power load prediction plug-ins;

[0160] P power load prediction branches are randomly selected from the K power load prediction branches of the J power load prediction plug-ins for prediction, and J initial load prediction intervals are output, wherein the initial load prediction interval is constructed based on P prediction results of the P power load prediction branches.

[0161] Wherein, the load fluctuation analysis is performed, and the number P of branch calls is set according to the load fluctuation coefficient, comprising:

[0162] The load fluctuation analysis is performed, and the load fluctuation coefficient is output according to the historical power load sequence, wherein the load fluctuation coefficient is the ratio of the standard deviation of the power load to the mean value in the historical power load sequence;

[0163] The load fluctuation coefficient is taken as a prediction complexity, and the ratio of the prediction complexity to the historical maximum prediction complexity in a historical time range is taken as a branch selection adjustment coefficient;

[0164] The product of the branch selection adjustment coefficient and the initial branch number is rounded to P, wherein the initial branch number is 5, and P is an integer greater than or equal to 3 and less than or equal to K.

[0165] Further, the power load prediction interval fitting module 15 comprises the following execution steps:

[0166] If J is 1, the initial load prediction interval is set as the power load prediction interval of the target power grid in the preset time zone;

[0167] If J is not 1, according to the interval fusion strategy configured based on the plurality of burst scenario similarities, the J initial load prediction intervals are fitted, and the power load prediction interval of the target power grid in the preset time zone is output.

[0168] According to the interval fusion strategy configured based on the plurality of burst scenario similarities, the J initial load prediction intervals are fitted, and the power load prediction interval of the target power grid in the preset time zone is output, comprising:

[0169] The J burst scenario similarities corresponding to the J initial load prediction intervals in the plurality of burst scenario similarities are obtained;

[0170] J fusion confidence weights are configured according to the J burst scenario similarities, wherein the fusion confidence weight is the ratio of the burst scenario similarity to the sum of the J burst scenario similarities;

[0171] The corresponding random sampling number is set according to the J fusion confidence weights, wherein the random sampling number is the product of the fusion confidence weight and a preset sampling number, and the preset sampling number is 1000;

[0172] Based on the corresponding random sampling number, the J initial load prediction intervals are fitted, and the power load prediction interval is output.

[0173] Based on the corresponding random sampling number, the J initial load prediction intervals are fitted, and the power load prediction interval is output, comprising:

[0174] Based on the corresponding random sampling number, the J initial load prediction intervals are fitted, and the power load prediction interval is output, comprising:

[0175] The minimum sampling power load data and the maximum sampling power load data in the sampling power load data set are selected to construct the power load prediction interval and output.

[0176] It should be noted that in the above embodiments, the description of each embodiment focuses on different aspects, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0177] Those skilled in the art will appreciate that embodiments of the application can be a method, a system or a computer program product. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer-readable program code.

[0178] The application is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more functions specified in the flowchart and / or block diagram. Figure 1 one or more functions specified in the flowchart and / or block diagram.

[0179] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more functions specified in the flowchart and / or block diagram. Figure 1 one or more functions specified in the flowchart and / or block diagram.

[0180] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more functions specified in the flowchart and / or block diagram. Figure 1 one or more functions specified in the flowchart and / or block diagram.

[0181] Although the preferred embodiments of the application have been described, those skilled in the art can make additional changes and modifications to the embodiments once they know the basic inventive concept.

[0182] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover modifications and variations of this application provided they come within the scope of the application and their equivalent technology.

Claims

1. A multi-objective optimization method for predicting power load intervals, characterized in that, The methods include: Based on the power grid operation monitoring records of the target power grid within a historical time range, feature extraction and feature clustering of sudden events are performed to determine several high-frequency sudden event feature data intervals; Several power load prediction plug-ins are constructed based on the aforementioned high-frequency burst characteristic data intervals, and combined to obtain a power load prediction model; The real-time sudden event characteristic data of the current power grid sudden event is compared with the several high-frequency sudden event characteristic data intervals to obtain the similarity of multiple sudden events. Based on the model invocation strategy configured according to the similarity of the multiple sudden scenarios, the power load prediction model is invoked to predict the power load within the preset time zone, and J initial load prediction intervals are output, where J is an integer greater than or equal to 1; Based on the similarity of the multiple sudden scenarios, an interval fusion strategy is configured to perform interval fitting on the J initial load prediction intervals, and the power load prediction interval of the target power grid within the preset time zone is output. Among them, several power load forecasting plugins are constructed based on the aforementioned high-frequency burst characteristic data intervals, including: A first high-frequency burst characteristic data interval is randomly selected. Using the first high-frequency burst characteristic data interval as a condition constraint, a sample power load sequence set is collected based on the power grid operation monitoring records of the target power grid within the historical time range. The historical power load of different sample power load sequences in the future historical time zone is obtained to obtain a sample predicted power load set. The sample power load sequence set and the sample predicted power load set are used as sample training data and divided into K equal parts. The first training set is constructed by selecting the first set with replacement K times. The first training set is constructed by iteratively selecting the first set K times, where K is an integer greater than or equal to 10 and less than or equal to 50. Using the K training sets, the Long Short-Term Memory Network is trained to convergence to obtain K first power load prediction branches, which are then integrated to construct the first power load prediction plugin. The plugins corresponding to the several high-frequency burst feature data intervals are then trained sequentially. The model invocation strategy is configured based on the similarity of the multiple sudden scenarios, including: If the number of similarities among the multiple sudden scenarios that are greater than or equal to the first similarity threshold is not zero, then the power load prediction plugin corresponding to the maximum similarity of the sudden scenario is selected as the adaptation prediction model. If the number of similarities of the multiple sudden scenarios greater than or equal to the first similarity threshold is 0 and the number of similarities of the multiple sudden scenarios greater than or equal to the second similarity threshold is not 0, then multiple power load prediction plug-ins corresponding to the multiple sudden scenario similarities greater than or equal to the second similarity threshold are selected as the adaptation prediction model, wherein the second similarity threshold is less than the first similarity threshold. If the number of similarities among the multiple sudden scenarios that are greater than or equal to the second similarity threshold is 0, then all power load prediction plug-ins in the power load prediction model are selected as the adapted prediction model.

2. The multi-objective optimization method for predicting power load intervals according to claim 1, characterized in that, Based on the power grid operation monitoring records of the target power grid within a historical time range, feature extraction and clustering of sudden events are performed to determine several high-frequency sudden event feature data intervals, including: Based on the power grid operation monitoring records of the target power grid within a historical time range, the characteristics of sudden events are extracted to obtain a sample sudden event feature dataset. The sudden event features include at least the weather type, weather data, power load fluctuation amplitude, equipment failure frequency, voltage fluctuation amplitude, power system frequency fluctuation amplitude, power grid load growth rate, instantaneous load change rate, and load distribution uniformity coefficient. The K-means clustering algorithm is used to cluster the sample burst feature dataset to determine several cluster centers and several high-frequency burst feature data intervals.

3. The multi-objective optimization method for predicting power load intervals according to claim 1, characterized in that, Perform power load forecasting within a preset time zone, and output J initial load forecast intervals, including: Collect historical power load sequences up to the current time point, perform load fluctuation analysis, and set the number of branch calls P based on the load fluctuation coefficient; The adaptive prediction model is invoked, wherein the adaptive prediction model contains J power load prediction plugins; Within the K power load prediction branches of the J power load prediction plugins, P power load prediction branches are randomly selected for prediction, and J initial load prediction intervals are output. The initial load prediction intervals are constructed based on the P prediction results of the P power load prediction branches.

4. The multi-objective optimization method for predicting power load intervals according to claim 3, characterized in that, Perform load fluctuation analysis and set the number of branch calls P based on the load fluctuation coefficient, including: The load fluctuation is calculated based on the historical power load sequence, and the load fluctuation coefficient is output. The load fluctuation coefficient is the ratio of the standard deviation and the mean of the power load in the historical power load sequence. The load fluctuation coefficient is used as the prediction complexity, and the ratio of the prediction complexity to the historical maximum prediction complexity within the historical time range is set as the branch selection adjustment coefficient. The product of the branch selection adjustment coefficient and the initial number of branches is rounded down to P, where the initial number of branches is 5, and P is an integer greater than or equal to 3 and less than or equal to K.

5. The multi-objective optimization method for predicting power load intervals according to claim 3, characterized in that, If J is 1, then the initial load forecast interval is set as the power load forecast interval of the target power grid within the preset time zone; If J is not 1, the interval fusion strategy is configured according to the similarity of the multiple sudden scenarios, and the interval fitting is performed on the J initial load prediction intervals to output the power load prediction interval of the target power grid in the preset time zone.

6. The multi-objective optimization method for predicting power load intervals according to claim 5, characterized in that, Based on the similarity configuration of the multiple sudden scenarios, an interval fusion strategy is configured to perform interval fitting on the J initial load prediction intervals, and output the power load prediction interval of the target power grid within the preset time zone, including: Obtain the J initial load prediction intervals corresponding to the J sudden scenario similarities in the multiple sudden scenario similarities; J fusion trust weights are configured based on the similarity of the J sudden scenarios, where the fusion trust weight is the ratio of the sudden scenario similarity to the sum of the similarities of the J sudden scenarios; The corresponding random sampling number is set according to the J fusion trust weights, wherein the random sampling number is the product of the fusion trust weights and the preset sampling number, wherein the preset sampling number is 1000; Based on the corresponding number of random samples, interval fitting is performed on the J initial load prediction intervals to output the power load prediction interval.

7. The multi-objective optimization method for predicting power load intervals according to claim 6, characterized in that, Based on the corresponding number of random samples, interval fitting is performed on the J initial load prediction intervals to output the power load prediction intervals, including: Based on the corresponding number of random samples, random sampling with replacement is performed in each of the J initial load prediction intervals to obtain a sampled power load dataset. The minimum and maximum sampled power load data in the sampled power load dataset are selected to construct and output the power load prediction interval.

8. A multi-objective optimization power load interval prediction system, characterized in that, A method for predicting power load intervals using multi-objective optimization as described in any one of claims 1 to 7, comprising: The sudden event feature extraction module is used to extract and cluster sudden event features based on the power grid operation monitoring records of the target power grid within a historical time range, and to determine several high-frequency sudden event feature data intervals. The power load forecasting model acquisition module is used to construct several power load forecasting plug-ins based on the aforementioned several high-frequency burst characteristic data intervals, and combine them to obtain a power load forecasting model; The emergency scene similarity acquisition module is used to compare the real-time emergency feature data of the current power grid emergency with the several high-frequency emergency feature data intervals to obtain the similarity of multiple emergency scenes. The power load forecasting model retrieval module is used to configure the model invocation strategy according to the similarity of the multiple sudden scenarios, invoke the power load forecasting model to perform power load forecasting within a preset time zone, and output J initial load forecast intervals, where J is an integer greater than or equal to 1; The power load forecast interval fitting module is used to configure an interval fusion strategy based on the similarity of the multiple sudden scenarios, perform interval fitting on the J initial load forecast intervals, and output the power load forecast interval of the target power grid within the preset time zone.

Images