Power supply output prediction method and device based on neural architecture search

Abstract

The invention provides a power output prediction method and device based on neural architecture search, and belongs to the field of power systems. The method comprises the following steps: acquiring natural condition data of a target power supply in a to-be-predicted time period; obtaining historical natural condition data having strong similarity with the natural condition data, and obtaining historical output data corresponding to the historical natural condition data; obtaining model input data according to the natural condition data, the historical natural condition data and the historical output data, and inputting the model input data into an output prediction model corresponding to the target power supply to obtain predicted output data of the target power supply in the to-be-predicted time period; wherein the output prediction model is determined by performing architecture search in a predefined super network based on training data of the target power supply. By using strong similar historical data and self-adaptive model architecture search, the output data of the target power supply in the to-be-predicted time period can be predicted more accurately, and the prediction precision and effectiveness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power systems, and in particular to a power output prediction method and device based on neural architecture search. BACKGROUND

[0002] Under the background of accelerating the transformation of energy structure to low carbonization, large-scale new energy grid connection brings great challenges to the stability and security of power grid operation: new energy power generation has significant intermittency and volatility characteristics, and its output has uncertainty. There is a contradiction between this uncertainty and the demand of power grid for stable and reliable power supply. When the proportion of new energy power generation reaches a certain level, the increase of new energy output prediction error will lead to the increase of power grid balancing cost, and even may trigger frequency limit accident, which seriously threatens the safe and stable operation of power grid. Therefore, it is urgent to improve the accuracy of new energy output prediction and accelerate the construction of clean, low-carbon, safe and efficient energy system.

[0003] In related technologies, the radiation transmission process is solved by relying on atmospheric physical equation, and new energy output prediction is performed from the physical mechanism level. However, the accuracy thereof depends on high-precision component parameters and data, and the accuracy is difficult to guarantee. Moreover, in actual application, the data acquisition cost is high, the efficiency is low, and the application has great limitations. SUMMARY

[0004] The embodiments of the present application provide a power output prediction method and device based on neural architecture search, to solve the problem of difficult efficient and accurate power output prediction.

[0005] In a first aspect, the embodiments of the present application provide a power output prediction method based on neural architecture search, comprising:

[0006] obtaining natural condition data of a target power source in a to-be-predicted time period;

[0007] obtaining historical natural condition data having strong similarity with the natural condition data, and obtaining historical output data corresponding to the historical natural condition data; wherein the strong similarity represents that the similarity degree is greater than a preset similarity threshold;

[0008] obtaining model input data according to the natural condition data, the historical natural condition data and the historical output data, inputting the model input data into an output prediction model corresponding to the target power source, and obtaining predicted output data of the target power source in the to-be-predicted time period output by the output prediction model; wherein the output prediction model is determined by architecture search in a pre-defined super network based on training data of the target power source.

[0009] In a possible implementation, before the model input data is input into the output prediction model corresponding to the target power source, the method further includes:

[0010] obtaining training data of the target power source and real output data corresponding to the training data;

[0011] taking the training data as input data and the real output data as labels, searching and optimizing the network architecture in a pre-defined super network until a difference between predicted output data of the obtained network architecture and corresponding real output data is less than a preset loss threshold, and obtaining an optimal network architecture;

[0012] constructing the output prediction model according to the optimal network architecture.

[0013] In a possible implementation, the training data includes first training data and second training data.

[0014] The taking the training data as input data and the real output data as labels, searching and optimizing the network architecture in a pre-defined super network includes:

[0015] alternately performing the first step and the second step;

[0016] the first step: keeping weights of connection operations of each node in the network architecture unchanged, searching and optimizing weights inside each connection operation based on the first training data and real output data corresponding to the first training data;

[0017] the second step: keeping the weights inside each connection operation unchanged, searching and optimizing the weights of the connection operations of each node based on the second training data and real output data corresponding to the second training data.

[0018] In a possible implementation, the obtaining natural condition data of the target power source in a to-be-predicted time period includes:

[0019] obtaining weather data of the to-be-predicted time period;

[0020] performing importance screening on the weather data to obtain natural condition data having strong correlation with output data, where the strong correlation represents a correlation degree greater than a preset correlation threshold.

[0021] In a possible implementation, the obtaining model input data according to the natural condition data, historical natural condition data and historical output data includes:

[0022] performing modal decomposition on the historical output data to obtain sub-modes of different central frequencies;

[0023] According to the similarity between the sub-modes, each sub-mode is classified and reorganized to obtain a plurality of new sub-modes; wherein, the new sub-modes include at least one sub-mode, and each sub-module in the new sub-modes has similar complexity and irregularity;

[0024] The natural condition data, the historical natural condition data, and each new sub-mode are used as model input data.

[0025] In a possible implementation, the new sub-modes are obtained by classifying and reorganizing each sub-mode according to the similarity between the sub-modes, and the method comprises the following steps of:

[0026] The fuzzy entropy of each sub-mode is calculated.

[0027] The sub-modes with a similarity greater than a preset similarity are combined to obtain new sub-modes, and the remaining sub-modes are used as new sub-modes respectively.

[0028] In a possible implementation, the method of obtaining the predicted output data of the target power source in the to-be-predicted time period output by the output prediction model comprises the following steps of:

[0029] The natural condition data, the historical natural condition data, and each new sub-mode are input into the output prediction model corresponding to the target power source to obtain sub-prediction data corresponding to each new sub-mode output by the output prediction model;

[0030] Each sub-prediction data is reconstructed to obtain the predicted output data.

[0031] In a possible implementation, after obtaining the predicted output data of the target power source in the to-be-predicted time period output by the output prediction model, the method further comprises the following steps of:

[0032] Load information of the power system is obtained.

[0033] According to the load information and the predicted output data, the charging and discharging of the energy storage system is optimized to balance the supply and demand of the power system.

[0034] In a second aspect, an embodiment of the present application provides a power output prediction device based on neural architecture search, comprising:

[0035] A first obtaining unit is configured to obtain natural condition data of a target power source in a to-be-predicted time period.

[0036] The second acquisition unit is configured to obtain historical natural condition data having strong similarity with the natural condition data, and obtain historical output data corresponding to the historical natural condition data; wherein the strong similarity represents that the degree of similarity is greater than a preset similarity threshold.

[0037] The first processing unit is configured to obtain model input data according to the natural condition data, the historical natural condition data, and the historical output data, input the model input data into an output prediction model corresponding to the target power source, and obtain predicted output data of the target power source in a to-be-predicted time period output by the output prediction model; wherein the output prediction model is determined by performing architecture search in a pre-defined super network based on training data of the target power source.

[0038] In a possible implementation manner, before the model input data is input into the output prediction model corresponding to the target power source, the apparatus further includes:

[0039] The third acquisition unit is configured to acquire training data of the target power source and real output data corresponding to the training data.

[0040] The execution unit is configured to search and optimize a network architecture in a pre-defined super network by taking the training data as input data and the real output data as a label, until a difference between predicted output data of the obtained network architecture and corresponding real output data is less than a preset loss threshold, and obtain an optimal network architecture.

[0041] The second processing unit is configured to construct the output prediction model according to the optimal network architecture.

[0042] The embodiment of the present application provides a power output prediction method and device based on neural architecture search, natural condition data of a target power source in a to-be-predicted time period is acquired; historical natural condition data with strong similarity with the natural condition data is obtained, and historical output data corresponding to the historical natural condition data is obtained; wherein the strong similarity represents that the similarity degree is greater than a preset similarity threshold; model input data is obtained according to the natural condition data, the historical natural condition data and the historical output data, and the model input data is input into an output prediction model corresponding to the target power source, to obtain predicted output data of the target power source in the to-be-predicted time period output by the output prediction model; wherein the output prediction model is determined by architecture search in a pre-defined super network based on training data of the target power source. By acquiring the natural condition data and screening the strong similar historical data and corresponding historical output data, more relevant and effective input information can be provided for prediction, the model input data is more suitable for the actual situation, and therefore the accuracy of prediction is improved; and the output prediction model is determined by architecture search based on the super network, the suitable model architecture can be automatically searched according to the characteristics of the target power source, the model is better fitted to the output law of the target power source, and the prediction precision is further improved. In summary, by using the strong similar historical data and adaptive model architecture search, the output data of the target power source in the to-be-predicted time period can be more accurately predicted, and the prediction precision and effectiveness are improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0044] Figure 1 is an implementation flowchart of a power output prediction method based on neural architecture search provided by the embodiment of the present application;

[0045] Figure 2 is an implementation flowchart of another power output prediction method based on neural architecture search provided by the embodiment of the present application;

[0046] Figure 3 is a structural schematic diagram of a power output prediction device based on neural architecture search provided by the embodiment of the present application. DETAILED DESCRIPTION

[0047] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known methods, devices, circuits, and SUMMARY

[0049] Under the background of accelerating the transformation of energy structure to low carbonization, large-scale photovoltaic grid-connected will bring great challenges to the stable operation and safe and economic operation of the power grid: research shows that the increase of photovoltaic output prediction error will lead to the rise of power grid balancing cost, and in extreme weather such as typhoon and sandstorm, prediction deviation may even trigger frequency out-of-limit accident. Therefore, it is urgent to improve the accuracy of photovoltaic output prediction and accelerate the construction of clean, low-carbon, safe and efficient energy system.

[0050] In the related art, the time series model based on statistical learning is used for photovoltaic output prediction, the autocorrelation characteristics of the output sequence are mined through historical data, but it is difficult to model the dynamic coupling relationship between meteorological conditions and output, resulting in large prediction error in mutation weather scene; the numerical weather prediction driven model based on physical mechanism is used for photovoltaic output prediction, relying on atmospheric physical equation to solve the radiation transmission process, although it has physical interpretability, it depends on high-precision component parameters and minute-level meteorological data, and there are significant bottlenecks in engineering deployment.

[0051] In order to improve the accuracy and efficiency of power output prediction, in the embodiments of the present application, by acquiring natural condition data and screening strongly similar historical data and corresponding historical output data, more relevant and effective input information can be provided for prediction, so that the model input data is more in line with the actual situation, thereby improving the accuracy of prediction; and the output prediction model is determined based on super network architecture search, which can automatically search for suitable model architecture according to the characteristics of the target power source, so that the model can better fit the output law of the target power source, and further improve the prediction accuracy.

[0052] In order to make the objects, technical solutions and advantages of the present application clearer, specific embodiments will be described below with reference to the accompanying drawings.

[0053] Figure 1 The implementation flowchart of the power output prediction method based on neural architecture search provided by the embodiments of the present application is described in detail as follows:

[0054] In step 101, natural condition data of the target power source in the to-be-predicted time period is acquired.

[0055] Exemplarily, the range of the to-be-predicted time period is determined according to the predicted demand, for example, 1 hour, 1 day, or 1 week in the future, and the natural condition data of the corresponding time period is obtained.

[0056] In this embodiment, the natural condition data is obtained by a meteorological monitoring station, satellite remote sensing, a sensor network, and the like, and includes, but is not limited to, illumination intensity, temperature, humidity, wind speed, air pressure, and the like.

[0057] In a feasible implementation, the original data obtained is cleaned to remove outliers and missing values, and is standardized or normalized to ensure data quality and consistency.

[0058] In step 102, historical natural condition data having strong similarity with the natural condition data is obtained, and historical output data corresponding to the historical natural condition data is obtained; wherein the strong similarity represents that the degree of similarity is greater than a preset similarity threshold.

[0059] Exemplarily, the similarity between the natural condition data of the to-be-predicted time period and the historical data is calculated by using any one or more of the Euclidean distance, cosine similarity, dynamic time warping, and the like, and the historical natural condition data having a similarity greater than a preset similarity threshold (for example, a similarity greater than 0.8) is obtained as the strong similarity data. The filtered historical natural condition data is associated with the corresponding historical output data to form a data pair.

[0060] In this embodiment, the dimensions are weighted to reflect the influence of different factors on the power output of the power source.

[0061] In a feasible implementation, this embodiment preferentially selects historical data close to the to-be-predicted time period to better reflect the current power source characteristics and environmental change trend.

[0062] In step 103, model input data is obtained according to the natural condition data, the historical natural condition data, and the historical output data, and the model input data is input into the output prediction model corresponding to the target power source to obtain the predicted output data of the target power source in the to-be-predicted time period output by the output prediction model; wherein the output prediction model is determined by architecture search in a pre-defined super network based on the training data of the target power source.

[0063] Exemplarily, the natural condition data to be predicted is combined with the filtered historical natural condition data and the historical output data to construct a multi-dimensional feature vector as the model input data, and the constructed model input data is input into the trained model to obtain the predicted output data of the target power source in the to-be-predicted time period.

[0064] In a feasible implementation, the embodiment performs post-processing on the predicted output data, such as confidence interval estimation, outlier correction, etc., to improve the reliability of the prediction.

[0065] In a feasible implementation, the embodiment defines an hyper-network in advance, which contains multiple candidate network structures (such as different numbers of layers, different numbers of neurons, and different connection modes). The hyper-network is used to search for an optimal network architecture corresponding to the target power source by using reinforcement learning, genetic algorithm, or gradient descent method, and the training data of the target power source is used as the basis, and the minimization of prediction error is used as the target for optimization, so as to obtain an output prediction model corresponding to the target power source.

[0066] In summary, the embodiment of the application can provide more relevant and effective input information for prediction by obtaining natural condition data and screening historical data and corresponding historical output data that are strongly similar, so that the model input data is more in line with the actual situation, thereby improving the accuracy of the prediction. The output prediction model is determined based on the hyper-network architecture search, which can automatically search for a suitable model architecture according to the characteristics of the target power source, so that the model can better fit the output law of the target power source, thereby further improving the prediction accuracy.

[0067] Figure 2 Another implementation flowchart of the power output prediction method based on neural architecture search provided by the embodiment of the application is provided, and is described in detail as follows:

[0068] In step 201, natural condition data of a target power source in a to-be-predicted time period is obtained.

[0069] In a feasible implementation, the step includes the following steps:

[0070] In a feasible implementation, the step includes the following steps:

[0071] In a feasible implementation, the step includes the following steps:

[0072] In a feasible implementation, the step includes the following steps:

[0073] In step 202, historical natural condition data having strong similarity with the natural condition data is obtained, and historical output data corresponding to the historical natural condition data is obtained; the strong similarity represents that the degree of similarity is greater than a preset similarity threshold.

[0074] For example, the embodiment uses the Euclidean distance, cosine similarity and other methods to quantify the similarity between the to-be-predicted data and the historical data, performs weighted calculation on multi-dimensional parameters such as light and temperature (for example, light × 0.5 + temperature × 0.3), sets a threshold (for example, similarity ≥ 0.85) to screen strong similar samples. The screened historical natural condition data and the output data at the corresponding moment are accurately matched, missing values are completed by interpolation, abnormal samples such as equipment failure are removed, and the data quality is ensured.

[0075] In step 203, model input data is obtained according to the natural condition data, the historical natural condition data and the historical output data.

[0076] In one example, step 203 includes the following steps:

[0077] The historical output data is subjected to modal decomposition to obtain a plurality of sub-modes with different center frequencies.

[0078] According to the similarity between the sub-modes, each sub-mode is classified and reorganized to obtain a plurality of new sub-modes; the new sub-modes include at least one sub-mode, and each sub-module in the new sub-modes has similar complexity and irregularity.

[0079] The natural condition data, the historical natural condition data and each new sub-mode are used as the model input data.

[0080] According to the similarity between the sub-modes, each sub-mode is classified and reorganized to obtain a new sub-mode, which can include:

[0081] The fuzzy entropy of each sub-mode is calculated; sub-modes with a similarity greater than a preset similarity are combined to obtain new sub-modes, and the remaining sub-modes are used as new sub-modes.

[0082] Exemplarily, the embodiment decomposes the historical output data into a plurality of intrinsic mode functions (IMFs) by using empirical mode decomposition or variational mode decomposition (VMD), each of which corresponds to a fluctuation characteristic of different frequency, such as long-term trend, seasonal change or random noise. The decomposition precision is controlled by setting the number of decomposition layers. According to the similarity of complexity and irregularity between the IMFs, the sub-modes are reorganized and classified, wherein the similarity of complexity and irregularity between the IMFs can be determined by the numerical distance of the fuzzy entropy corresponding to the IMFs, for example, the numerical distance is less than or equal to a first numerical value.

[0083] In one example, the embodiment quantifies the signal complexity by calculating the fuzzy entropy of each IMF. Based on the fuzzy entropy similarity, the sub-modes are clustered, the low-entropy IMFs (such as trend terms) with strong regularity and the high-entropy IMFs (such as noise terms) with high randomness are merged to form new sub-modes, and the reorganization and classification of the sub-modes are realized.

[0084] The embodiment aligns the natural condition data of the to-be-predicted time period, the historical similar condition data and the reorganized sub-modes according to time steps, and constructs a multi-dimensional feature matrix as the model input data.

[0085] The embodiment method realizes the multi-scale decomposition-clustering-fusion strategy, retains the complex fluctuation characteristics of the output data, reduces the dimension of the model input data, provides the prediction model with input data that has clear physical meaning and time sequence correlation, and can improve the prediction accuracy.

[0086] In step 204, training data of a target power supply and real output data corresponding to the training data are obtained.

[0087] Exemplarily, the embodiment obtains a plurality of training data of the target power supply and real output data corresponding to each training data, wherein each set of training data includes natural condition data, historical natural condition data similar to the natural condition data, and historical output data corresponding to the historical natural condition data.

[0088] In step 205, the training data are taken as input data, the real output data are taken as labels, and the network architecture is searched and optimized in a pre-defined super network until the difference between the output data predicted by the obtained network architecture and the corresponding real output data is less than a preset loss threshold, and the optimal network architecture is obtained.

[0089] Exemplarily, the embodiment presets a super network containing multiple connection operations, such as a long short-term memory (LSTM), a temporal convolutional network (TCN), a fully connected layer, and the like, defines a candidate architecture parameter space, supports dynamic combination to generate a subnetwork, adopts reinforcement learning, a genetic algorithm, or a differentiable search, takes loss minimization as an optimization objective, and iteratively trains the sub-architecture:

[0090] Randomly sample the sub-architecture in the super network, calculate the predicted output by forward propagation with the training data, and calculate the loss value with the real label. The search strategy is updated according to the loss value (such as the reward mechanism in reinforcement learning or the crossover and mutation probability in the genetic algorithm), and the architecture with a loss lower than the current average value is retained. When the validation loss (based on an independent validation set) of the optimal sub-architecture in the last N iterations (such as N = 10) is less than a preset threshold (such as mean square error ≤ 0.05), or the training round reaches an upper limit, the search is stopped, and the current optimal network architecture is output.

[0091] The process balances the model complexity and the prediction accuracy by automatically searching for a model structure that adapts to the characteristics of the target power supply, and avoids the limitations of manual parameter tuning.

[0092] In a feasible implementation, the training data includes first training data and second training data; and the step 205 includes the following steps:

[0093] The first step and the second step are alternately performed.

[0094] The first step: the weights of the connection operations of the nodes in the network architecture are kept unchanged, and the weights inside each connection operation are searched and optimized based on the first training data and the real output data corresponding to the first training data.

[0095] The second step: the weights inside each connection operation are kept unchanged, and the weights of the connection operations of the nodes are searched and optimized based on the second training data and the real output data corresponding to the second training data.

[0096] In one example, the first step of the embodiment (optimizing the internal weights of the connection operations): fixing the weights of the connection operations between nodes in the network architecture (such as whether the nodes are connected, the type of connection), only optimizing the internal parameters (such as the weights of the convolution kernel, the parameters of the LSTM gate) of the reserved connections. Using the first training data and its true output label, the internal weights of the connection are iteratively updated by optimizers such as Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (Adam), etc., so that the loss (such as mean square error) between the predicted output and the true value gradually decreases until convergence or a preset number of iterations is reached. This process focuses on mining the expression capacity of existing connections.

[0097] The second step (optimizing the weights of the connection operations between nodes): keeping the internal weights of the connections optimized in the first step unchanged, and adjusting the weights of the connection operations between nodes (such as attention mechanism weights, path selection probability). Based on the second training data and its label, reinforcement learning or discrete optimization methods (such as crossover and mutation in genetic algorithms) are used to dynamically select or generate node connection patterns (such as adding / deleting skip connections, adjusting inter-layer routing), and the second training data loss is used as a feedback signal to retain connection structure changes that can reduce the loss. This process focuses on exploring better network topology.

[0098] Alternating execution logic: through the alternating iteration of "fixed structure-optimized parameters-fixed parameters-optimized structure", the joint optimization of network architecture and parameters is realized.

[0099] In one possible implementation, the embodiment first constructs a super-network as a carrier for search, and the super-network is a parent network that contains all possible sub-network structures. Next, the characteristics of the super-network are introduced from the candidate operation set, mixed operation, and node connection mode.

[0100] The candidate operation set contains operations such as LSTM, Convolutional Neural Network (CNN), and Attention mechanism. Among them, CNN uses convolution to locally connect a part of the input and extract corresponding features. LSTM introduces the hidden node state h t-1 of the last time slice and the gate mechanism to control the flow and loss of features. The attention mechanism achieves the function of quickly extracting important information by assigning weights to features.

[0101] The basic unit of search is a directed acyclic graph composed of an ordered sequence of N nodes. Each node x (i) is a latent representation (for example, a feature map in a convolutional network), and each directed edge (i,j) is associated with a transformation x (i)Each intermediate node in a cell receives all possible connections from its predecessor nodes, and information fusion is achieved by weighted summation as follows:

[0102]

[0103] Discrete candidate operations are continuous, and each operation in the candidate operation set is applied to each node x (i) To make the search space continuous, the categorical selection of a specific operation is relaxed to a mixture of all possible operations as follows:

[0104]

[0105] where represents the candidate operation set, represents the mixed operation of the connection (i, j) node pair, and o(x) represents any discrete operation.

[0106] This embodiment adopts a gradient descent method to jointly learn the architecture a and the weight w within all mixed operations, and finally finds the optimal parameters a * . Use and to represent the loss on the first training data D train and the loss on the second training data D val . These two losses are not only determined by the architecture a, but also by the weights w in the network. And searching for the optimal a * is to find the a * that minimizes the validation loss , where the weights associated with the architecture are obtained by minimizing the training loss . That is, we need to solve such a double-layer optimization problem, a as the upper variable, and w as the lower variable, as follows:

[0107]

[0108] After obtaining the optimal architecture parameters a * , the continuous architecture is derived into a discrete architecture. For each intermediate node, 1 strongest connection is reserved, where the strength of the edge is defined as follows, and finally the optimal network architecture for a specific target power source is obtained:

[0109]

[0110] Step 206, according to the optimal network architecture, an output prediction model is constructed.

[0111] Exemplarily, after determining the optimal network architecture, this embodiment further defines the loss function and the optimizer configuration and other information to construct the output prediction model.

[0112] Step 207, input the model input data into the output prediction model corresponding to the target power source to obtain the predicted output data of the target power source in the to-be-predicted time period output by the output prediction model, wherein the output prediction model is determined based on the training data of the target power source in the pre-defined super network through architecture search.

[0113] In an implementable embodiment, step 207 comprises the following steps:

[0114] The natural condition data, historical natural condition data, and each new sub-mode are input into the output prediction model corresponding to the target power source to obtain the sub-prediction data corresponding to each new sub-mode output by the output prediction model.

[0115] Each sub-prediction data is reconstructed to obtain the predicted output data.

[0116] Exemplarily, the model of the embodiment inputs the natural condition data, historical natural condition data, and reorganized sub-modes in time sequence format into the output prediction model, respectively processes different features by using the multi-branch structure of the model, extracts time sequence features through convolution layers and cycle layers, and outputs the sub-prediction data corresponding to each sub-mode; then, according to the inverse process of the mode decomposition, the sub-prediction results of each sub-mode are stacked from low to high in frequency, or are combined by using the weight matrix learned through training (such as attention mechanism dynamically allocating the contribution degree of the sub-mode), and finally the complete output sequence of the to-be-predicted time period is restored, so that the reconstructed predicted data contains both long-term trend and short-term fluctuation characteristics.

[0117] Step 208, acquiring the load information of the power system; and optimizing the charge and discharge setting of the energy storage system according to the load information and the predicted output data to balance the supply and demand of the power system.

[0118] Exemplarily, the embodiment collects load data in real time, constructs a load prediction model combining historical load curves and meteorological factors (such as air temperature and holidays), acquires load prediction data of future time periods, and establishes an optimization model with the minimum curtailment rate, peak-valley difference flattening, or the lowest comprehensive cost as the target. For example, when the predicted output > load, the energy storage system is controlled to charge (absorb excess power); when the predicted output < load, the energy storage is discharged (supplement power gap) to balance the supply and demand of the power system.

[0119] In summary, the embodiment obtains weather data of a to-be-predicted time period through multi-source data fusion, retains strongly correlated variables through correlation screening, combines historical similar data matching and quality cleaning to provide high-correlation input for the model; through modal decomposition and fuzzy entropy clustering to reorganize sub-modes, the complexity of the output data is retained while the input dimension is reduced, the physical interpretability and time series correlation of the data features are improved; relying on the super network architecture search and alternating optimization strategy, the target power characteristics are automatically adapted, the model complexity and prediction accuracy are balanced, and the limitations of manual parameter tuning are avoided; finally, the model reconstructs the sub-prediction data, dynamically optimizes the energy storage charging and discharging strategy combined with the load information, forms a closed loop of prediction and regulation, effectively improves the power output prediction accuracy, and can provide data-driven scientific decision support for power supply and demand balance.

[0120] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the application.

[0121] The following is a device embodiment of the application, and for details not described in detail, reference can be made to the corresponding method embodiments described above.

[0122] Figure 3 The structure of the power output prediction device based on neural architecture search provided by the embodiment of the application is shown, only the part related to the embodiment of the application is shown for convenience of description, and the details are as follows:

[0123] As Figure 3 shown, the power output prediction device based on neural architecture search includes:

[0124] The first acquisition unit 31 is configured to acquire natural condition data of a target power source in a to-be-predicted time period.

[0125] The second acquisition unit 32 is configured to obtain historical natural condition data having strong similarity with the natural condition data, and obtain historical output data corresponding to the historical natural condition data; wherein the strong similarity represents that the similarity is greater than a preset similarity threshold.

[0126] The first processing unit 33 is configured to obtain model input data according to the natural condition data, the historical natural condition data and the historical output data, and input the model input data into an output prediction model corresponding to the target power source to obtain predicted output data of the target power source in the to-be-predicted time period output by the output prediction model; wherein the output prediction model is determined through architecture search in a pre-defined super network based on training data of the target power source.

[0127] In a feasible implementation manner, before the first processing unit 33, the device further includes:

[0128] The third obtaining unit is configured to obtain training data of the target power supply and real output data corresponding to the training data.

[0129] The execution unit is configured to search and optimize the network architecture in the predefined super network with the training data as input data and the real output data as labels, until a difference between output data predicted by the obtained network architecture and the corresponding real output data is less than a preset loss threshold, and an optimal network architecture is obtained.

[0130] The second processing unit is configured to construct an output prediction model according to the optimal network architecture.

[0131] In an implementation, the training data includes first training data and second training data.

[0132] The execution unit is specifically configured to: alternately perform the first step and the second step; wherein the first step: keep the weights of the connection operations of the nodes unchanged, and search and optimize the weights inside each connection operation based on the first training data and the real output data corresponding to the first training data. The second step: keep the weights inside each connection operation unchanged, and search and optimize the weights of the connection operations of the nodes based on the second training data and the real output data corresponding to the second training data.

[0133] In an implementation, the first obtaining unit 31 is specifically configured to:

[0134] Obtain weather data of a to-be-predicted time period.

[0135] Perform importance screening on the weather data to obtain natural condition data having strong correlation with the output data; wherein the strong correlation represents that the degree of correlation is greater than a preset correlation threshold.

[0136] In an implementation, the first processing unit 33 is specifically configured to:

[0137] Perform modal decomposition on the historical output data to obtain a plurality of sub-modes with different center frequencies.

[0138] According to the similarity between the sub-modes, classify and recombine each sub-mode to obtain a plurality of new sub-modes; wherein the new sub-modes include at least one sub-mode, and each sub-module in the new sub-modes has similar complexity and irregularity.

[0139] Take the natural condition data, the historical natural condition data, and each new sub-mode as model input data.

[0140] In an implementation, the first processing unit 33 is specifically configured to:

[0141] Calculate the fuzzy entropy of each sub-mode.

[0142] Merge the sub-modes with a similarity greater than a preset similarity between the fuzzy entropies to obtain new sub-modes, and take the remaining sub-modes as new sub-modes respectively.

[0143] In a feasible implementation, the first processing unit 33 is specifically further configured to:

[0144] Input the natural condition data, the historical natural condition data, and each new sub-mode into an output prediction model corresponding to the target power source to obtain sub-prediction data corresponding to each new sub-mode output by the output prediction model.

[0145] Reconstruct each sub-prediction data to obtain predicted output data.

[0146] In a feasible implementation, after the first processing unit 33, the device further includes an optimization unit specifically configured to: acquire load information of the power system; and optimize the charge-discharge setting of the energy storage system according to the load information and the predicted output data to balance the supply and demand of the power system.

[0147] In the above embodiments, the description of each embodiment has its own focus, and the parts not described or recorded in detail in a certain embodiment can be referred to the related description of other embodiments.

[0148] Those skilled in the art can appreciate that the templates, units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0149] If the modules / units are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware. The computer program can be stored in a computer readable storage medium, and when executed by a processor, the computer program can implement the steps of each power output prediction method based on neural architecture search. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0150] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those ordinarily skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A power output prediction method based on neural architecture search, characterized in that, The method comprises the following steps: obtaining natural condition data of a target power source in a to-be-predicted time period; obtaining historical natural condition data having strong similarity with the natural condition data, and obtaining historical output data corresponding to the historical natural condition data; wherein the strong similarity represents that the degree of similarity is greater than a preset similarity threshold; inputting model input data obtained according to the natural condition data, the historical natural condition data and the historical output data into an output prediction model corresponding to the target power source to obtain predicted output data of the target power source in the to-be-predicted time period output by the output prediction model; wherein the output prediction model is determined by architecture search in a pre-defined super network based on training data of the target power source. 2.The power output prediction method based on neural architecture search according to claim 1, wherein, Before the model input data is input into the output prediction model corresponding to the target power source, the method further comprises the following steps: obtaining training data of the target power source and real output data corresponding to the training data; searching and optimizing a network architecture in a pre-defined super network with the training data as input data and the real output data as labels until the difference between the output data predicted by the obtained network architecture and the corresponding real output data is less than a preset loss threshold, thereby obtaining an optimal network architecture; constructing the output prediction model according to the optimal network architecture. 3.The power output prediction method based on neural architecture search according to claim 2, wherein, The training data comprises first training data and second training data; the searching and optimizing of the network architecture in the pre-defined super network with the training data as input data and the real output data as labels comprises the following steps: alternately performing a first step and a second step; the first step: keeping the weights of connection operations in each node unchanged, searching and optimizing the weights inside each connection operation based on the first training data and the real output data corresponding to the first training data; the second step: keeping the weights inside each connection operation unchanged, searching and optimizing the weights of the connection operations of each node based on the second training data and the real output data corresponding to the second training data.

4. The neural architecture search based power output prediction method according to any one of claims 1-3, characterized in that, The obtaining of the natural condition data of the target power source in the to-be-predicted time period comprises the following steps: obtaining weather data in the to-be-predicted time period; performing importance screening on the weather data to obtain natural condition data having strong correlation with output data; wherein the strong correlation represents that the degree of correlation is greater than a preset correlation threshold.

5. The neural architecture search based power output prediction method according to any one of claims 1-3, wherein, The obtaining of the model input data according to the natural condition data, the historical natural condition data and the historical output data comprises the following steps: performing modal decomposition on the historical output data to obtain sub-modes of different central frequencies; classifying and reorganizing each sub-mode according to the similarity between the sub-modes to obtain new sub-modes; wherein the new sub-modes comprise at least one sub-mode, and each sub-module in the new sub-modes has similar complexity and irregularity; taking the natural condition data, the historical natural condition data and each new sub-mode as model input data. 6.The power output prediction method based on neural architecture search according to claim 5, wherein, The classifying and reorganizing of each sub-mode according to the similarity between the sub-modes comprises: calculating the fuzzy entropy of each sub-mode; merging the sub-modes with a similarity greater than a preset similarity, to obtain new sub-modes, and taking the remaining sub-modes as new sub-modes. 7.The power output prediction method based on neural architecture search according to claim 5, wherein, The method comprises: inputting the natural condition data, the historical natural condition data, and each new sub-mode into the output prediction model corresponding to the target power source, to obtain sub-prediction data corresponding to each new sub-mode output by the output prediction model; reconstructing each sub-prediction data to obtain the predicted output data.

8. The neural architecture search based power output prediction method according to any one of claims 1-3, wherein, After obtaining the predicted output data of the target power source in the to-be-predicted time period output by the output prediction model, the method further comprises: obtaining load information of a power system; optimizing the charge-discharge setting of an energy storage system according to the load information and the predicted output data, to balance the supply and demand of the power system. 9.A power output prediction device based on neural architecture search, characterized by, The method comprises: a first obtaining unit configured to obtain natural condition data of a target power source in a to-be-predicted time period; a second obtaining unit configured to obtain historical natural condition data having strong similarity with the natural condition data, and obtain historical output data corresponding to the historical natural condition data; wherein the strong similarity represents a similarity greater than a preset similarity threshold; a first processing unit configured to obtain model input data according to the natural condition data, the historical natural condition data, and the historical output data, and input the model input data into an output prediction model corresponding to the target power source, to obtain predicted output data of the target power source in the to-be-predicted time period output by the output prediction model; wherein the output prediction model is determined by architecture search in a pre-defined hypernet based on training data of the target power source.

10. The neural architecture search based power output prediction device of claim 9, wherein, Before the model input data is input into the output prediction model corresponding to the target power source, the apparatus further comprises: a third obtaining unit configured to obtain training data of the target power source and real output data corresponding to the training data; a performing unit configured to search and optimize a network architecture in a pre-defined hypernet with the training data as input data and the real output data as a label, until the difference between the output data predicted by the obtained network architecture and the corresponding real output data is less than a preset loss threshold, to obtain an optimal network architecture; a second processing unit configured to construct the output prediction model according to the optimal network architecture.

Images