Electricity forecasting method and system based on electric power index analysis

Abstract

The present application provides a power prediction method and system based on power index analysis. During network debugging, the basic power state detection network can imitate the feature processing performance of the target control network, and the basic power state detection network can imitate the characteristic value distribution of the target control network. Then, the target power state detection network generated by the representation vector extraction module in the debugged target power state classification network can accurately extract the power state representation vector, and the target power state detection network does not require redundant training data during the establishment process, which reduces the cost of network debugging and improves efficiency. Furthermore, because the target power state classification network is obtained through transfer learning debugging, the target power state classification network can realize the data analysis capabilities of complex large models through a sophisticated small model, which reduces the hardware environment requirements for the deployed equipment and saves computing power and deployment costs.

Description

Technical Field

[0001] The present application relates to the fields of data processing, artificial intelligence, and power analysis, and specifically, to a power forecasting method and system based on power index analysis. Background Art

[0002] In rural construction, reasonable planning and safe and continuous output of electric energy are the prerequisites for ensuring stability. This is inseparable from the power status analysis, comprehensive evaluation and prediction of each rural power system. How to conduct an accurate, low-cost and highly adaptable power status analysis system is a technical issue that fits the rural power maintenance. Summary of the invention

[0003] The purpose of this application is to provide a power forecasting method and system based on power index analysis.

[0004] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present application.

[0005] According to one aspect of an embodiment of the present application, there is provided a power forecasting method based on power index analysis, comprising:

[0006] Acquire a large power system data set collected from the target power system;

[0007] Using the debugged target power state detection network, extracting the power state representation vector of the power system big data set, and obtaining the quasi-paired power state representation vector corresponding to the target power system;

[0008] Determine the power state of the target power system by using the characteristic distance between each power state characterization vector deployed in advance and the power state characterization vector to be paired;

[0009] The debugging process of the target power status detection network includes:

[0010] Acquire a power data learning sample set; the power data learning sample set includes a set of power data learning sample sets collected from each example power system and power state training indication information corresponding to each example power system;

[0011] Through the power data learning sample set, the power state classification reference network is repeatedly debugged multiple times to obtain a debugged target control network, and the target cluster centroids generated by the target control network for each of the example power systems are determined;

[0012] Determine the centroids of each target cluster as the centroids of each cluster maintained in the basic power state classification network, and perform multiple transfer learning debugging on the basic power state classification network based on the power data learning sample set and the target control network to obtain a debugged target power state classification network;

[0013] A target power state detection network is generated through a target representation vector extraction module in the target power state classification network.

[0014] In one embodiment, determining the power state of the target power system by using the characteristic distance between each pre-deployed power state characterization vector and the power state characterization vector to be paired includes:

[0015] Acquire each power state characterization vector deployed in advance, and respectively determine the characteristic distance between the power state characterization vector to be paired and each of the power state characterization vectors;

[0016] When it is determined that there is a target characteristic distance that meets the predetermined requirements among the characteristic distances, the power state corresponding to the target characteristic distance that meets the set requirements is used as the power state of the target power system.

[0017] In one embodiment, the power state classification reference network includes a characterization vector extraction reference module and a classification reference module. When the power state classification reference network is repeatedly debugged, it includes:

[0018] Loading the acquired first power data learning sample into the representation vector extraction reference module to obtain a reference power state representation vector corresponding to the first power data learning sample;

[0019] Determine, by means of the classification reference module, reference space distances between the reference power state representation vector and the centroids of each currently generated inference cluster, and determine classification reference information estimated for the first power data learning sample based on each reference space distance;

[0020] Optimizing the network learnable variables of the power state classification reference network and the centroids of each inference cluster by using the error between the classification reference information and the power state training indication information corresponding to the first power data learning sample;

[0021] The method of determining the centroids of each target cluster as the centroids of each cluster maintained in the basic power state classification network, and performing multiple transfer learning debugging on the basic power state classification network based on the power data learning sample set and the target control network to obtain the debugged target power state classification network includes:

[0022] Determining each target cluster centroid as each cluster centroid maintained in the basic power state classification network, and repeatedly debugging the basic power state classification network based on the power data learning sample set to obtain a debugged process power state classification network;

[0023] The process power state classification network is subjected to multiple transfer learning debugging through the power data learning sample set and the target control network to obtain a debugged target power state classification network.

[0024] In one embodiment, the basic power state classification network includes a basic characterization vector extraction module and a basic decision module. When the basic power state classification network is repeatedly debugged, it includes:

[0025] Loading the acquired second power data learning sample into the basic representation vector extraction module to obtain a first basic power state representation vector corresponding to the second power data learning sample;

[0026] Loading the first basic power state representation vector into the basic decision module, obtaining a first basic decision result obtained by reasoning the first basic power state representation vector and the centroid vectors of each target cluster;

[0027] The network learnable variables of the basic power state classification network are optimized by using the error between the first basic decision result and the power state training indication information corresponding to the second power data learning sample.

[0028] In one embodiment, the first basic decision result obtained by reasoning the first basic power state representation vector and the centroid vectors of each target cluster includes:

[0029] Determining, by means of the basic decision module, first basic spatial distances between the first basic power state representation vector and the centroids of each target cluster;

[0030] A first basic decision result estimated for the second power data learning example is determined through each first basic spatial distance.

[0031] In one embodiment, the process power state classification network includes a process representation vector extraction module and a process decision module. When a round of transfer learning debugging is performed on the process power state classification network, it includes:

[0032] Loading the acquired third power data learning sample into the target control network to obtain a first reference power state representation vector extracted by the target control network;

[0033] Loading the third power data learning sample into the debugged process representation vector extraction module to obtain a corresponding process power state representation vector, and obtaining an estimated process decision result based on the process decision module through the process power state representation vector and the centroids of each target cluster;

[0034] The network learnable variables of the process power state classification network are optimized by the error between the first reference power state characterization vector and the process power state characterization vector, and by the error between the process decision result and the power state training indication information corresponding to the third power data learning sample.

[0035] In one embodiment, the optimizing the network learnable variables of the process power state classification network by the error between the first reference power state characterization vector and the process power state characterization vector, and by the error between the process decision result and the power state training indication information corresponding to the third power data learning example, comprises:

[0036] Calculating a first transfer learning error result by an error between the first reference power state characterization vector and the process power state characterization vector, and calculating a first classification error result by an error between the process decision result and the power state training indication information corresponding to the third power data learning example;

[0037] Network learnable variables of the process power state classification network are optimized through the first transfer learning error result and the second classification error result.

[0038] In one embodiment, when performing a round of transfer learning debugging on the basic power state classification network, it includes:

[0039] Loading the acquired fourth power data learning sample into the target control network to obtain a second reference power state representation vector extracted by the target control network, and loading the fourth power data learning sample into the basic power state classification network to obtain an extracted second basic power state representation vector and an estimated second basic decision result;

[0040] The network learnable variables of the basic power state classification network are optimized by the error between the second reference power state characterization vector and the second basic power state characterization vector, and by the error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning sample.

[0041] In one embodiment, the basic power state classification network includes a basic characterization vector extraction module and a basic decision module, and the fourth power data learning sample is loaded into the basic power state classification network to obtain the extracted second basic power state characterization vector and the estimated second basic decision result, including:

[0042] Loading the fourth power data learning sample into the basic representation vector extraction module to obtain an extracted second basic power state representation vector;

[0043] Obtaining an estimated second basic decision result through the basic decision module, through the second basic power state characterization vector and the centroids of each target cluster;

[0044] The optimizing the network learnable variables of the basic power state classification network by the error between the second reference power state characterization vector and the second basic power state characterization vector, and by the error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning example, comprises:

[0045] Calculate a second transfer learning error result based on an error between the second reference power state characterization vector and the second basic power state characterization vector, and calculate a second classification error result based on an error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning example;

[0046] The network learnable variables of the initial power state classification network to be debugged are optimized through the second transfer learning error result and the second classification error result.

[0047] According to another aspect of an embodiment of the present application, there is provided a power prediction system, including:

[0048] processor;

[0049] and a memory for storing executable instructions for the processor;

[0050] The processor is configured to perform the above method by executing the executable instructions.

[0051] This application has at least the following beneficial effects:

[0052] The power prediction method and system based on power index analysis provided in the embodiment of the present application can apply the centroids of each cluster generated by the debugged target control network to the basic power state classification network by determining the centroids of each cluster maintained in the basic power state detection network, and maintain them unchanged during the debugging of the basic power state classification network, so that when the debugging is completed, the stable point of the basic power state representation vector extraction network is close to the stable point of the target control network. At the same time, relying on the target control network and the power data learning sample set, the basic power state detection network can be transferred and learned, which not only allows the basic power state detection network to imitate the target control network during debugging, but also can effectively prevent the basic power state detection network from being damaged. According to the feature processing performance of the network, and the basic power state detection network can imitate the characteristic value distribution of the target control network, then the target power state detection network generated by the representation vector extraction module in the debugged target power state classification network can accurately extract the power state representation vector, and the target power state detection network does not require redundant training data in the establishment stage, which reduces the cost of network debugging and improves efficiency. Furthermore, because the target power state classification network is obtained through transfer learning debugging, the target power state classification network can realize the data analysis capabilities of complex large models through sophisticated small models, reducing the hardware environment requirements for deployed equipment and saving computing power and deployment costs.

[0053] It should be understood that the above general description and the following detailed description are exemplary and explanatory only and are not intended to limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present application, and together with the specification are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0055] Figure 1 It is a flow chart of a power forecasting method based on power index analysis provided in an embodiment of the present application.

[0056] Figure 2 It is a schematic diagram of a process of performing power status detection based on a target power status detection network provided in an embodiment of the present application.

[0057] Figure 3 It is a schematic diagram of the functional module architecture of the power prediction device provided in an embodiment of the present application.

[0058] Figure 4 It is a schematic diagram of the composition of a power prediction system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as being limited to the examples set forth herein; on the contrary, these embodiments are provided so that the present application will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art. In addition, the described features, structures, or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical solutions of the present application may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations, or operations are not shown or described in detail to avoid blurring the various aspects of the present application.

[0060] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0061] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0062] When analyzing the power state (such as power load state, power supply capacity state, market operation state, carbon emission state, supply and demand balance results, etc.) based on the power index, it is necessary to rely on the power state detection network (trained and debugged deep learning neural network) to extract the power state characterization vector (vector information that characterizes the power state characteristics). For the power state detection network, for both cost and performance considerations, it is expected that the number of network parameters of the power state detection network is small to reduce the dependence on computing power. At the same time, it is expected that the power state detection network has excellent power state characterization vector extraction accuracy and efficiency. It is conceivable that the above two expected dimensions are contradictory to each other. In the debugging stage of the power state detection network, the reference power state detection network with a large number of network parameters is used to limit the power state detection network to be debugged with a small number of network parameters, or based on the quantification of the power state detection network with a large number of network parameters, a power state detection network with a small number of parameters is obtained. However, for the power state detection network with a large number of network parameters, the power state detection network to be debugged with a small number of network parameters is restricted. The power state detection network is restricted in turn on the feature graphs of each level. The rigid restriction is likely to cause insufficient generalization performance of the network, resulting in increased debugging costs. In addition, the number of costs restricted at different levels is different, and the computational cost of obtaining the cost is also a considerable overhead, which not only affects the debugging efficiency of the network, but also reduces the performance of the debugged network. On the other hand, for the power state detection network with a small number of parameters obtained by quantizing the power state detection network with a large number of network parameters, although the power state detection network with a large number of network parameters can improve the accuracy of the power state representation vector extraction, when the quantization is completed, the detection accuracy of the network is likely to decrease, and more training data is required for quantitative debugging, which increases the training cost.

[0063] The embodiment of the present application provides a power prediction method and system based on power index analysis. In the debugging phase of the network, after obtaining a power data learning sample set, the power state classification reference network is repeatedly debugged multiple times through the power data learning sample set to obtain a debugged target control network, and the centroids of each target cluster generated inside the target control network are obtained, and then the centroids of each target cluster are determined as the centroids of each cluster maintained in the basic power state classification network, and based on the target control network and the aforementioned power data learning sample set, the basic power state classification network is repeatedly transferred and debugged to obtain a debugged target power state classification network, and then a target power state detection network is generated through a representation vector extraction module in the target power state classification network. Based on this, the target cluster centroids generated by the debugged target control network are determined as the cluster centroids maintained in the basic power state detection network. The cluster centroids generated by the target control network can be applied to the basic power state classification network, and maintained unchanged during the debugging of the basic power state classification network. When the debugging is completed, the stable point of the basic power state representation vector extraction network is close to the stable point of the target control network. At the same time, relying on the target control network and the power data learning sample set, the basic power state detection network can be transferred for learning, which not only allows the basic power state detection network to imitate the feature processing performance of the target control network during debugging, but also If the basic power state detection network can imitate the eigenvalue distribution of the target control network, then the target power state detection network generated by the representation vector extraction module in the debugged target power state classification network can accurately extract the power state representation vector, and the target power state detection network does not require redundant training data during the establishment phase, which reduces the cost of network debugging and improves efficiency. Furthermore, because the target power state classification network is obtained through transfer learning debugging, the target power state classification network can realize the data analysis capabilities of complex large models through a sophisticated small model, reducing the hardware environment requirements for the deployed equipment and saving computing power and deployment costs.

[0064] The following is a detailed description of the debugging process of the target power status detection network, which includes the following steps:

[0065] Step S110, obtaining a power data learning sample set.

[0066] In the process of debugging and optimizing the power state detection network in the embodiment of the present application, the classification network including the characterization vector extraction module is debugged, and finally the target characterization vector extraction module obtained by debugging is used to generate the target power state detection network. In order to meet the debugging requirements of the classification network, the embodiment of the present application first determines each power state category, and collects the power data learning samples corresponding to each category, such as collecting the power big data of the power system corresponding to the power state. As the power data learning sample, the power data can include multiple power indexes, specifically including power consumption index, power supply index, market index, and dual carbon index. The electricity consumption index reflects the electricity consumption of various industries, such as the electricity consumption of the primary industry, the electricity consumption of urban and rural residents, and the electricity consumption of key enterprises; the power supply index reflects the power supply capacity of the power supply unit, such as the power generation of local power plants, the power generation of centralized power plants, and the power generation of local power grids; the market index reflects the trading volume and average transaction price of different types of electricity in the market, such as the trading volume of each trading variety and the average transaction price of each trading variety; the dual-carbon index reflects the construction of green and low-carbon, such as clean energy power generation, clean energy transmission volume, and electricity substitution. The concept of electricity index has been disclosed in the relevant technology and will not be elaborated here. The above-mentioned types of power data are sorted to obtain power data learning samples, that is, debugging samples for training. The power status training indication information corresponding to each example power system, wherein the power status training indication information of the example power system is the supervision information of the debugging link, for example, it can be a label marking the power status of the corresponding example power system.

[0067] After determining each example power system, the power data learning samples generated by each example power system are collected respectively, and a subset of power data learning samples corresponding to each example power system is constructed respectively through the power data learning samples corresponding to each example power system, so as to obtain a match between the subset of power data learning samples and the power state training indication information (label and other supervision information) of the corresponding example power system. In this way, the obtained power data learning sample set includes the subset of power data learning samples collected for each example power system and the power state training indication information corresponding to each example power system.

[0068] Step S120, repeatedly debugging the power state classification reference network through the power data learning sample set to obtain a debugged target control network, and determine the target cluster centroid generated by the target control network for each example power system.

[0069] After obtaining the power data learning sample set, the power state classification reference network is repeatedly debugged through the power data learning sample set to obtain the debugged target control network, wherein the target control network is a neural network obtained after the debugging of the power state classification reference network meets the preset first debugging stop requirement (i.e., the evaluation requirement of network convergence, such as reaching the maximum number of debugging times, the error reaching the minimum, etc.). The power state classification reference network provided in the embodiment of the present application includes a representation vector extraction reference module and a classification reference module. The representation vector extraction reference module can be a convolutional neural network (including a filter kernel, an activation function, and a pooling kernel) for extracting information vectors representing power state features; after obtaining the power state representation vector extracted by the representation vector extraction reference module, the classification reference module obtains the reference space distance (the distance represents the distance between two vectors, such as the Euclidean distance, the cosine distance, etc., which can represent the similarity between the two) through the power state representation vector and the generated cluster centroid (i.e., the center of each class obtained by clustering), and then based on the classifier (such as the Softmax of multi-classification), the estimated decision result is obtained according to the reference space distance, and the decision result is the classification information obtained by classification.

[0070] In the embodiment of the present application, when debugging to obtain the target control network, step S120 may specifically include:

[0071] Step S121: extracting a first power data learning sample from a power data learning sample set.

[0072] When the power state classification reference network is repeatedly debugged (i.e., one round of iterative optimization is completed), the number of power data learning samples loaded simultaneously in a Batch is used to extract the corresponding number of power data learning samples from the power data learning sample set and determine them as an input sample, that is, the first power data learning sample used in one round of debugging.

[0073] Step S122: Load the acquired first power data learning sample into a representation vector extraction reference module to obtain a reference power state representation vector corresponding to the first power data learning sample.

[0074] For example, after obtaining the first power data learning sample, the first power data learning sample is loaded into a representation vector extraction reference module (for example, a convolutional neural network) in a power state classification reference network to obtain a reference power state representation vector output by the representation vector extraction reference module through the first power data learning sample. When the input first power data learning sample includes multiple power data learning samples, the representation vector extraction reference module extracts the power state representation vector for each power data learning sample, obtains the reference power state multi-dimensional feature information corresponding to each power data learning sample, and refers to the reference power state representation vector corresponding to the first power data learning sample.

[0075] Step S123, through the classification reference module, respectively determine the reference space distances between the reference power state representation vector and the centroids of each currently generated inference cluster, and determine the classification reference information estimated for the first power data learning sample based on each reference space distance.

[0076] After obtaining the reference power state representation vector extracted by the representation vector extraction reference module, the reference power state representation vector is loaded into the classification reference module, and based on the classification reference module, the reference space distances between the reference power state representation vector and the centroids of each currently generated inference cluster are determined respectively, and then through each reference space distance, projection (i.e., vector mapping) is used to obtain the classification reference information estimated for the first power data learning sample.

[0077] For example, in the classification reference module, a corresponding inference cluster centroid is generated for each state classification. In the embodiment of the present application, the generated inference cluster centroids correspond to the reference codes of each example power system (such as 001), wherein the inference cluster centroid is P (dimension) × Q (number of example power systems), and each inference cluster centroid is in a learnable state in the debugging link of the power state classification reference network. Then, the current inference cluster centroids are respectively subjected to matrix multiplication operations with the extracted reference power state representation vectors to obtain the reference space distance, and a softmax classifier is used to determine the classification reference information estimated for the first power data learning sample based on the obtained reference space distance. When there are multiple power data learning samples in the first power data learning sample loaded in a Batch, the obtained classification reference information is the confidence or probability parameter of each power data learning sample corresponding to each state classification.

[0078] Step S124 , optimizing the network learnable variables and the centroids of each inference cluster of the power state classification reference network by using the error between the classification reference information and the power state training indication information corresponding to the first power data learning example.

[0079] After obtaining the classification reference information corresponding to the first power data learning sample, the loss is determined based on the determined loss function (such as cross entropy) through the error between the classification reference information and the power state training indication information corresponding to the first power data learning sample, and the network learnable variables of the power state classification reference network are optimized based on the loss, and the centroids of each inference cluster generated by the power state classification reference network are optimized through the loss. In the embodiment of the present application, when there are multiple power data learning samples in the first power data learning sample, for each power data learning sample, the loss is calculated based on the loss function through the error between the classification reference information corresponding to the power data learning sample and the corresponding power state training indication information, so as to optimize (such as general optimization methods such as gradient descent algorithm, adaptive algorithm, etc.) the network learnable variables (such as weights, bias, learning rate, etc. parameters) and each inference cluster centroid (predicted cluster centroid) of the power state classification reference network through the weighted results of the calculated losses.

[0080] Step S125, if the first debugging stop requirement is met, execute S126, if not, return to execute S121.

[0081] When the current repeated debugging is completed, the embodiment of the present application classifies the debugging status of the reference network through the current power status to analyze whether it meets the first debugging stop requirement. A repeated debugging is a complete reverse transfer through a batch of first power data learning samples.

[0082] Step S126, outputting the debugged target control network.

[0083] Executing step S126 represents the convergence of the reference power state detection network. At this time, the power state classification reference network obtained in the current debugging can be determined as the debugged target control network, and the centroids of each inference cluster generated in the power state classification reference network obtained in the current debugging can be determined as the centroids of the target clusters generated for each example power system.

[0084] Based on this, by debugging the power state classification reference network (i.e., a complex large model) with a large number of network parameters, we can not only obtain the target control network for state classification through power data, but also obtain the target cluster centroid with high robustness generated inside the target control network, which helps provide a reference standard for the debugging of the basic power state classification network.

[0085] The following describes the steps of debugging the power status classification reference network, including the various functional modules (software or hardware) involved and their relationship with each step:

[0086] For debugging of the power state classification reference network, a debugging data providing module, a power state characterization vector extraction module, a cluster centroid preservation module, a loss determination module and an optimization adjustment module are set.

[0087] Among them, the debugging data providing module is used to extract the first power data learning sample from the power data learning sample set during the debugging phase, generate a batch of the extracted first power data learning sample to load it into the representation vector extraction reference module of the power state classification reference network, which can be used to execute step S121.

[0088] The power state characterization vector extraction module is used to perform state feature mining on the power data learning sample through the characterization vector extraction reference module to obtain a vector that can characterize the power state, that is, to obtain a reference power state characterization vector, which can be used to execute step S122.

[0089] The cluster centroid saving module is used to save the inference cluster centroid generated by the power state classification reference network for each state classification, providing a reference for step S123. The number of state classifications of the power state classification reference network is pre-configured according to actual needs.

[0090] The loss determination module is used to perform matrix multiplication operations on the reference power state representation vector extracted by the representation vector extraction reference module and the centroid of each inference cluster to obtain the reference space distance, and project the reference space distance to the probability domain according to the classifier to obtain the probability of each state classification corresponding to each power data learning sample to obtain the classification reference information, and then obtain the error (loss) through the error between the classification reference information and the power state training indication information corresponding to the first power data learning sample to execute step S123.

[0091] The optimization and adjustment module is used to optimize and adjust the learnable variables of the power state classification reference network, and optimize the centroids of each inference cluster generated to execute step S124.

[0092] Step S130, determining each target cluster centroid as each cluster centroid maintained in the basic power state classification network, and performing multiple transfer learning debugging on the basic power state classification network based on the power data learning sample set and the target control network to obtain a debugged target power state classification network.

[0093] After debugging and obtaining the target control network and the target cluster centroids generated by the target control network, the target cluster centroids are determined as the cluster centroids maintained in the basic (i.e., the initial) power state classification network. The maintained cluster centroids represent that when debugging the basic power state classification network, the cluster centroids are fixed. In other words, the state of the cluster centroids in the basic power state classification network is determined to be unlearnable. Then, the basic power state classification network is repeatedly debugged through the target control network and the power data learning sample set, and the debugged target power state classification network is obtained when the preset first debugging stop requirement is met.

[0094] When debugging the basic power state classification network, different debugging links may be adopted, and there may be the following two ways to debug the basic power state classification network. The basic power state classification network includes a basic representation vector extraction module and a basic decision module. The basic representation vector extraction module is, for example, a convolutional neural network, and the basic decision module is, for example, a Softmax classifier. As mentioned above, the two debugging methods for debugging the basic power state classification network are specifically:

[0095] The first method is: based on the training data set, by maintaining the basic power state classification network with the centroid of each cluster, the process power state classification network (i.e., an intermediate module, a transition module) is debugged, and then the process power state classification network is transferred and debugged through the power data learning sample set and the target control network to obtain the debugged target power state classification network.

[0096] In the first mode, when the error in the number of network parameters between the target control network and the basic power state classification network is large, in order to prevent the basic power state classification network from being forced to learn the network performance of the target control network, causing the basic power state classification network to overfit, the embodiment of the present application performs multi-stage debugging on the basic power state classification network in the debugging of the first mode. Specifically, the basic power state classification network that maintains the centroid of each target cluster is firstly debugged, so that the power state representation vector extracted by the basic power state classification network gradually converges around the centroid of each target cluster, so as to learn the distribution of the characteristic values ​​of the target control network and obtain the process power state classification network, and then, based on transfer learning (also known as knowledge distillation), the process power state classification network is debugged by transfer learning, and the error of the power state representation vector extracted by the target control network and the process power state classification network, as well as the error of the process power state classification network, is combined to fine-tune the process power state classification network to obtain the target power state classification network.

[0097] When debugging the target power state classification network, the debugging of the target power state classification network may specifically include:

[0098] Step S131, determining the centroids of each target cluster as the centroids of each cluster maintained in the basic power state classification network, and repeatedly debugging the basic power state classification network based on the power data learning sample set to obtain a debugged process power state classification network.

[0099] Specifically, the target cluster centroids generated by the target control network are applied to the basic power state classification network, and the cluster centroids in the basic power state classification network are maintained without offset during debugging. The basic power state classification network with the cluster centroids set is repeatedly debugged through the power data learning sample set to obtain the debugged process power state classification network.

[0100] More specifically, the intermediate process of debugging the process power status classification network includes:

[0101] Step S1311: extracting a second power data learning sample from the power data learning sample set.

[0102] When the basic power state classification network is repeatedly debugged, the number of power data learning samples loaded simultaneously in a Batch is used to extract the corresponding number of power data learning samples from the power data learning sample set and determine them as one input, that is, the second power data learning sample used for one debugging.

[0103] Step S1312: Load the acquired second power data learning sample into the basic representation vector extraction module to obtain the first basic power state representation vector corresponding to the second power data learning sample.

[0104] After obtaining the second power data learning sample, the obtained second power data learning sample is loaded into the basic representation vector extraction module in the basic power state classification network to obtain the basic power state representation vector output by the basic representation vector extraction module through the second power data learning sample. When there are multiple power data learning samples in the loaded second power data learning sample, the basic representation vector extraction module extracts the power state representation vector for each power data learning sample, so as to obtain the basic power state representation vectors corresponding to each power data learning sample in a batch, and the initial power state multi-dimensional feature information corresponding to each power data learning sample in a batch is regarded as the basic power state representation vector corresponding to the second power data learning sample.

[0105] Step S1313: Load the first basic power state characterization vector into the basic decision module to obtain a first basic decision result obtained by reasoning the first basic power state characterization vector and the centroid vectors of each target cluster.

[0106] After obtaining the first basic power state characterization vector extracted by the basic characterization vector extraction module, the first basic power state characterization vector is loaded into the basic decision module, and based on the basic decision module, the first basic spatial distances between the first basic power state characterization vector and the current centroids of each target cluster are determined respectively, and the first basic decision result of the first power data learning sample is projected through each first basic spatial distance.

[0107] The embodiment of the present application determines the first basic spatial distance between the first basic power state characterization vector and each target cluster centroid respectively through the basic decision module, and then determines the first basic decision result of the second power data learning sample estimation through each first basic spatial distance. In the basic decision module, the first basic power state characterization vector is matrix multiplied with each target cluster centroid respectively to obtain each first basic spatial distance, and then the first basic decision result of the second power data learning sample estimation is determined based on each first basic spatial distance through the classifier. When there are multiple power data learning samples in a batch of loaded second power data learning samples, the first basic decision result obtained is the probability of each power data learning sample corresponding to each state classification. In this way, relying on the basic decision module, the corresponding decision result can be determined based on the maintained centroids of each target cluster and the extracted first basic power state characterization vector, so as to gradually optimize and adjust in the debugging link of the basic power state classification network, and help the extracted power state characterization vector to be distributed in the centroids of each target cluster.

[0108] Step S1314, optimizing the network learnable variables of the basic power state classification network by using the error between the first basic decision result and the power state training indication information corresponding to the second power data learning sample.

[0109] After obtaining the first basic decision result corresponding to the second power data learning sample, the error between the first basic decision result and the power state training indication information corresponding to the second power data learning sample is used to calculate the loss through a preset loss function, and the network learnable variables of the basic power state classification network are optimized based on the loss.

[0110] Step S1315, when the second debugging stop requirement is met, execute S1316, if not met, execute S1311.

[0111] Step S1316, outputting the debugged process power status classification network.

[0112] Step S132, performing multiple transfer learning debugging on the process power state classification network through the power data learning sample set and the target control network to obtain a debugged target power state classification network.

[0113] After debugging the process power state classification network, the process representation vector extraction module and the process decision module included in the process power state classification network can be obtained. The process representation vector extraction module is obtained by debugging the basic representation vector extraction module in the basic power state classification network, and the process decision module is obtained by debugging the basic decision module in the basic power state classification network.

[0114] The following describes the process of performing multiple transfer learning and debugging on the process power state classification network through the power data learning sample set and the target control network to obtain the debugged target power state classification network, including:

[0115] Step 1321: extract a third power data learning sample from the power data learning sample set.

[0116] The present application can determine the number of power data learning samples in a batch loaded into the process power state classification network during the first generation debugging of the process power state classification network according to actual conditions, and determine the corresponding number of power data learning samples obtained from the power data learning sample set as the third power data learning samples.

[0117] Step S1322: Load the acquired third power data learning sample into the target control network to obtain a first reference power state representation vector extracted by the target control network.

[0118] For example, the acquired third power data learning sample is loaded into the target control network to obtain the first reference power state representation vector output by the target representation vector extraction reference module in the target control network. The target representation vector extraction reference module is debugged by the representation vector extraction reference module in the power state classification reference network.

[0119] Step S1323, load the third power data learning sample into the debugged process characterization vector extraction module to obtain the corresponding process power state characterization vector, and based on the process decision module, obtain the estimated process decision result through the process power state characterization vector and the centroid of each target cluster.

[0120] For example, the third power data learning sample is loaded into the process power state classification network, and based on the process characterization vector extraction module in the process power state classification network, the corresponding process power state characterization vector is obtained, and the process power state characterization vector is loaded into the process decision module in the process power state classification network. Through the process decision module, the estimated process decision result is obtained through the obtained process power state characterization vector and the centroid of each target cluster.

[0121] Step S1324, optimizing the network learnable variables of the process power state classification network through the error between the first reference power state characterization vector and the process power state characterization vector, and through the error between the process decision result and the power state training indication information corresponding to the third power data learning example.

[0122] A first transfer learning error result is calculated by the error between the first reference power state characterization vector and the process power state characterization vector, and a first classification error result is calculated by the error between the process decision result and the power state training indication information corresponding to the third power data learning sample. Then, the network learnable variables of the process power state classification network are optimized through the first transfer learning error result and the second classification error result.

[0123] The embodiment of the present application uses a target control network to perform a quadratic constraint on the power state characterization vector extracted by the process power state classification network to limit the features extracted by the process power state classification network. At the same time, because the number of parameters of the target control network is greater and the expression effect is better, the power state characterization vector can be detected more accurately through the target control network. Based on this, by performing a quadratic constraint on the target control network, the process power state classification network can converge to the same eigenvalue distribution as the target control network again, thereby increasing the detection accuracy of the process power state classification network.

[0124] The loss function for determining the first transfer learning error result by the error between the first reference power state characterization vector and the process power state characterization vector may be a Euclidean distance loss function.

[0125] When calculating the first classification error result through the error between the process decision result and the power state training indication information corresponding to the third power data learning sample, the first classification error result between the process decision result and the power state training indication information corresponding to the third power data learning sample is determined through the cross entropy loss function.

[0126] If the third power data learning sample includes x power data learning samples, then when determining the first classification error result, the first classification branch error between the process decision result obtained through the power data learning sample and the power state training indication information corresponding to the power data learning sample is determined for each of the x power data learning samples, and each first classification branch error is weighted averaged, and the result is determined as the first classification error result calculated for the third power data learning sample. After obtaining the first transfer learning error result and the first classification error result, in order to optimize the process power state classification network so that the process power state classification network can learn the eigenvalue distribution of the target control network, the first transfer learning error result and the first classification error result are collaboratively optimized, for example:

[0127] C=α·C 11 +C 12

[0128] Among them, C is the coordinated error, α is the adjustment coefficient (set according to actual needs), C 11 is the first transfer learning error result, C 12is the first classification error result.

[0129] The network learnable variables in the process power state classification network are optimized and adjusted through collaborative errors, and the centroids of each cluster of the process power state classification network and the target control network are kept unchanged.

[0130] Under the first debugging method, when debugging the basic power state classification network and the process power state classification network, the size of the learning rate is set according to actual needs for optimization. Since the process power state classification network is debugged to obtain the target power state classification network, it is a fine adjustment based on the process power state classification network. The learning rate can be set relatively smaller to debug the process power state classification network. For example, when debugging the basic power state classification network, rate=0.1, and when debugging the process power state classification network, rate=0.01.

[0131] Based on this, by determining the first transfer learning error result and the first classification error result, when optimizing the process power state classification network, the feature extraction error and classification error between the process power state classification network and the target control network are taken into account at the same time, and the obtained target power state classification network can accurately complete the detection of the power state characterization vector.

[0132] Step S1325, when the preset third debugging stop requirement is met, execute S1326, and when it is not met, return to execute S1321.

[0133] Step S1326, outputting the debugged target power state classification network.

[0134] In this way, when conducting transfer learning and completing distillation debugging, the eigenvalue distribution between the process power state classification network and the target control network can be synchronously limited based on the collaborative error limit, as well as the decision results of the process power state classification network, to complete the re-optimization of the process power state classification network, so as to obtain the target power state classification network that can accurately extract the power state representation vector and complete the power state classification based on this.

[0135] Furthermore, the debugging method based on the first method can prevent the forced intervention of the basic power state classification network to learn the performance of the target control network when the number of network parameters between the basic power state classification network and the target control network is greatly different, thereby causing network overfitting. At the same time, by debugging the process of obtaining the target power state classification network, there is no need to add redundant data during network debugging. The basic power state classification network can be debugged in stages based on the general debugging method to obtain a target power state classification network with excellent detection performance. The target power state detection network with a small number of network parameters can achieve the performance of the target control network with a large number of network parameters. In other words, the performance of a complex large model is achieved with a sophisticated small model, reducing dependence on the hardware environment, and obtaining a target power state classification network with high stability and generalization.

[0136] When debugging the basic power state classification network in this application, each step is executed by a functional module (software or hardware is not limited). For the debugging of the basic power state classification network, a debugging data providing module, a first power state characterization vector extraction module, a first loss determination module, a second power state characterization vector extraction module, a second loss determination module, a cluster centroid storage module and a collaborative optimization adjustment module are set. Specifically:

[0137] The debugging data providing module is used to extract power data learning samples from the power data learning sample set during the debugging phase, and to organize the extracted power data learning samples into a batch of data and load it into the basic representation vector extraction module of the basic power state classification network, or load it into the process representation vector extraction module of the process power state classification network, and step S1321 can be executed.

[0138] The first power state characterization vector extraction module is used to mine the power state characteristics of the second power data learning sample through the basic characterization vector extraction module in different debugging stages, and to mine the power state characteristics of the third power data learning sample based on the process characterization vector extraction module. The power state characterization vectors of different debugging stages obtained have the power state distribution information of the power data learning sample, and step S1322 can be executed.

[0139] The first loss determination module is used to obtain the error between the estimated decision result and the corresponding power state training indication information in the unused debugging stage, and the matrix multiplication result of the basic power state characterization vector and the centroid of each cluster is projected through the basic decision module to obtain the probability (for example, a vector) of the second power data learning sample corresponding to each power state, which is determined as the basic decision result, and the loss is determined based on the error between the obtained basic decision result and the corresponding power state training indication information. In addition, the probability of the third power data learning sample corresponding to each power state obtained by the matrix multiplication projection of the process power state characterization vector and the centroid of each cluster can be determined as the process decision result through the process decision module, and the first classification error result can be determined based on the error between the obtained process decision result and the corresponding power state training indication information, and some steps of steps S1313, S3123 and S1324 can be executed.

[0140] The second power state representation vector extraction module is used to extract the representation vector for the third power data learning sample according to the target representation vector extraction reference module in the target control network, which can be used to execute step S1322.

[0141] The second loss determination module is used to obtain a first transfer learning error result by calculating the distance error between the first reference power state characterization vector extracted by the target control network and the process power state characterization vector extracted by the process characterization vector extraction module when debugging the process power state classification network, which can be used to execute step S1324.

[0142] The cluster centroid saving module is used to save the cluster centroids generated by each state classification in the basic power state classification network and the process power state classification network, which can provide a reference for the processing of steps S1313 and S1323. The number of state classifications of the basic power state classification network is set according to actual needs when generating the basic power state classification network.

[0143] The collaborative optimization adjustment module is used to optimize the network learnable variables (such as weights, biases, learning rates, etc.) through the obtained losses.

[0144] The second method is to debug the target control network and the power data learning sample set, perform multiple transfer learning debugging on the basic power state classification network that maintains the centroid of each cluster, and obtain the debugged target power state classification network.

[0145] If the number of network parameters between the target control network and the basic power state classification network is not much different, the target control network and the power data learning sample set can be used in the second method to perform multiple transfer learning debugging on the basic power state classification network that maintains the centroids of each cluster. The specific process is the same as the migration debugging in the first method. The following is an introduction to the process of debugging to obtain the target network, including:

[0146] Step S130a: extracting a fourth power data learning sample from the power data learning sample set.

[0147] For example, based on the actual debugging situation, the number of power data learning samples in a batch loaded into the basic power state classification network at the same time during a debugging phase of the basic power state classification network can be determined, and the corresponding number of power data learning samples obtained from the power data learning sample set can be determined as the fourth power data learning samples. In different generations of debugging the basic power state classification network, the fourth power data learning samples obtained are different.

[0148] Step S130b, load the acquired fourth power data learning sample into the target control network, obtain the second reference power state characterization vector extracted by the target control network, load the fourth power data learning sample into the basic power state classification network, obtain the extracted second basic power state characterization vector and the estimated second basic decision result.

[0149] For example, the acquired fourth power data learning sample is loaded into the target control network to obtain the second reference power state representation vector extracted by the target control network, and the fourth power data learning sample is loaded into the basic representation vector extraction module in the basic power state classification network to obtain the extracted second basic power state representation vector, and then according to the basic decision module in the basic power state classification network, the estimated second basic decision result is obtained through the second basic power state representation vector and the centroid of each target cluster.

[0150] The basic power state classification network includes a basic representation vector extraction module and a basic decision module. The structure of the basic representation vector extraction module is as mentioned above, and can be a convolutional neural network. The basic decision module is a classifier, such as Softmax, which performs matrix multiplication operations between the second basic power state representation vector and the centroid of each target cluster to obtain each second basic space distance, and performs feature clustering through each second basic space distance. The basic power state classification network is consistent with the target control network architecture, and the number of network parameters of the target control network is more than that of the basic power state classification network. In this way, relying on the structure of the basic power state classification network, the power state representation vector required for restriction and the estimated decision results can be obtained, which helps to learn and optimize the basic power state classification network.

[0151] Step S130c, optimizing the network learnable variables of the basic power state classification network through the error between the second reference power state characterization vector and the second basic power state characterization vector, and through the error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning sample.

[0152] For example, the second transfer learning error result is determined by the error between the second reference power state characterization vector and the second basic power state characterization vector, and the second classification error result is calculated based on the error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning sample. Then, the network learnable variables of the initial power state classification network to be debugged are optimized through the second transfer learning error result and the second classification error result.

[0153] The relevant principles have been introduced in the previous content and will not be repeated here.

[0154] Based on this, when the basic power state classification network is optimized by using the second transfer learning error result and the second classification error result obtained, the basic power state detection network is made to imitate the feature processing performance of the target control network during debugging, and the basic power state detection network is made to imitate the characteristic value distribution of the target control network, thereby helping the target power state detection network constructed by the representation vector extraction module in the debugged target power state classification network to accurately detect the power state representation vector.

[0155] Step S130d: If the preset fourth debugging stop requirement is met, execute step S130e; if not, continue to execute step S130a.

[0156] Step S130e, outputting the debugged target power state classification network.

[0157] Based on this, the target cluster centroids generated by the debugged target control network are determined as the cluster centroids maintained in the basic power state detection network. After the cluster centroids generated by the target control network are applied to the basic power state classification network, the transfer learning debugging based on the basic power state classification network allows the basic power state classification network to imitate the feature processing performance and eigenvalue distribution of the target control network. Then, the target power state detection network constructed by the representation vector extraction module in the debugged target power state classification network can accurately extract the power state representation vector.

[0158] Step S140, constructing a target power state detection network through a target representation vector extraction module in a target power state classification network.

[0159] In the embodiment of the present application, when the target power state classification network is obtained after debugging, the target power state detection network is constructed through the characterization vector extraction module in the target power state classification network, and the target characterization vector extraction module is obtained by debugging the basic characterization vector extraction module in the basic power state classification network. Based on this, after the target power state classification network is debugged, only the structure used for extracting the power state characterization vector in the target power state classification network is obtained. Then, for the target power state classification network obtained by debugging, the centroids of each target cluster and the decision module that the debugging link depends on are used to collaboratively complete the debugging of the characterization vector extraction module, so that the target power state detection network that accurately extracts the power state characterization vector can be obtained in the end.

[0160] Please refer to Figure 2 , is a flow chart of the target power state detection network obtained through the above debugging in the embodiment of the present application for power state detection, including the following steps:

[0161] Step S101, obtaining a power system big data set collected from a target power system.

[0162] Step S102 , extracting power state characterization vectors from the power system big data set through the debugged target power state detection network to obtain a quasi-paired power state characterization vector corresponding to the target power system.

[0163] Step S103 , determining the power state of the target power system through the characteristic distance between each pre-deployed power state characterization vector and the power state characterization vector to be paired.

[0164] For example, each pre-deployed power state characterization vector is obtained, and the characteristic distances between the power state characterization vector to be paired and each power state characterization vector are determined respectively; and when a target characteristic distance that meets predetermined requirements exists among the determined characteristic distances, the power state corresponding to the target characteristic distance that meets the set requirements is used as the power state of the target power system; if there is no target characteristic distance that meets the predetermined requirements (such as being greater than a preset threshold) among the characteristic distances, it means that the detection has failed, and the current power state is a new power state that can be added to the iteration of the network.

[0165] It should be noted that although the steps of the method in the present application are described in a specific order in the drawings, this does not require or imply that the steps must be performed in this specific order, or that all the steps shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps, etc.

[0166] The following introduces an embodiment of the device of the present application, which can be used to execute the power prediction method based on power index analysis in the above-mentioned embodiment of the present application. Figure 3 The structure block diagram of the power prediction device provided by the embodiment of the present application is schematically shown. Figure 3 As shown, the power prediction device 200 includes:

[0167] The data acquisition module 210 is used to acquire a large data set of the power system collected from the target power system;

[0168] The feature mining module 220 is used to extract the power state representation vector of the power system big data set by using the debugged target power state detection network to obtain the quasi-paired power state representation vector corresponding to the target power system;

[0169] A state detection module 230 is used to determine the power state of the target power system by using the characteristic distance between each power state characterization vector deployed in advance and the power state characterization vector to be paired;

[0170] The network debugging module 240 is used to debug the target power status detection network, specifically including:

[0171] Acquire a power data learning sample set; the power data learning sample set includes a set of power data learning sample sets collected from each example power system and power state training indication information corresponding to each example power system;

[0172] Through the power data learning sample set, the power state classification reference network is repeatedly debugged multiple times to obtain a debugged target control network, and the target cluster centroids generated by the target control network for each of the example power systems are determined;

[0173] Determine the centroids of each target cluster as the centroids of each cluster maintained in the basic power state classification network, and perform multiple transfer learning debugging on the basic power state classification network based on the power data learning sample set and the target control network to obtain a debugged target power state classification network;

[0174] A target power state detection network is generated through a target representation vector extraction module in the target power state classification network.

[0175] The specific details of the power prediction device provided in each embodiment of the present application have been described in detail in the corresponding method embodiments and will not be repeated here.

[0176] Figure 4 The structure block diagram of the computer system (ie, power prediction system) of the electronic device used to implement the embodiment of the present application is schematically shown.

[0177] It should be noted that Figure 4 The computer system 300 of the electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0178] like Figure 4 As shown, the computer system 300 includes a central processing unit 301 (CPU), which can perform various appropriate actions and processes according to the program stored in the read-only memory 302 (ROM) or the program loaded from the storage part 308 to the random access memory 303 (RAM). Various programs and data required for system operation are also stored in the random access memory 303. The central processing unit 301, the read-only memory 302 and the random access memory 303 are connected to each other through a bus 304. The input / output interface 305 (Input / Output interface, i.e., I / O interface) is also connected to the bus 304.

[0179] The following components are connected to the input / output interface 305: an input section 306 including a keyboard, a mouse, etc.; an output section 307 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card, a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the input / output interface 305 as needed. A removable medium 311, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 310 as needed so that a computer program read therefrom is installed into the storage section 308 as needed.

[0180] In particular, according to an embodiment of the present application, the process described in each method flow chart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flow chart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 309, and / or installed from the removable medium 311. When the computer program is executed by the central processing unit 301, various functions defined in the system of the present application are executed. It should be noted that the computer-readable medium shown in the embodiment of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections with one or more conductors, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device, or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, device, or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, etc., or any suitable combination of the foregoing.

[0181] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the above-mentioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0182] It should be noted that, although several modules or units of the equipment for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into being embodied by multiple modules or units.

[0183] Through the description of the above implementation methods, it is easy for those skilled in the art to understand that the example implementation methods described here can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the implementation methods of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation methods of the present application.

[0184] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary technical means in the art that are not disclosed in the present application.

[0185] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A power forecasting method based on power index analysis, characterized in that: include: Acquire a large power system data set collected from the target power system; Using the debugged target power state detection network, extracting the power state representation vector of the power system big data set, and obtaining the quasi-paired power state representation vector corresponding to the target power system; Determine the power state of the target power system by using the characteristic distance between each power state characterization vector deployed in advance and the power state characterization vector to be paired; The debugging process of the target power status detection network includes: Acquire a power data learning sample set; the power data learning sample set includes a set of power data learning sample sets collected from each example power system and power state training indication information corresponding to each example power system; Through the power data learning sample set, the power state classification reference network is repeatedly debugged multiple times to obtain a debugged target control network, and the target cluster centroids generated by the target control network for each example power system are determined, wherein the cluster centroids are the class centers obtained by clustering; The power state classification reference network includes a characterization vector extraction reference module and a classification reference module. When the power state classification reference network is repeatedly debugged, it includes: Loading the acquired first power data learning sample into the representation vector extraction reference module to obtain a reference power state representation vector corresponding to the first power data learning sample; Determine, by means of the classification reference module, reference space distances between the reference power state representation vector and the centroids of each currently generated inference cluster, and determine classification reference information estimated for the first power data learning sample based on each reference space distance; Optimizing the network learnable variables of the power state classification reference network and the centroids of each inference cluster by using the error between the classification reference information and the power state training indication information corresponding to the first power data learning sample; Determine the centroids of each target cluster as the centroids of each cluster maintained in the basic power state classification network, and perform multiple transfer learning debugging on the basic power state classification network based on the power data learning sample set and the target control network to obtain a debugged target power state classification network; A target power state detection network is generated through a target representation vector extraction module in the target power state classification network.

2. The method according to claim 1, characterized in that The determining the power state of the target power system by using the characteristic distance between each pre-deployed power state characterization vector and the power state characterization vector to be paired includes: Acquire each power state characterization vector deployed in advance, and respectively determine the characteristic distance between the power state characterization vector to be paired and each of the power state characterization vectors; When it is determined that there is a target characteristic distance that meets the predetermined requirements among the characteristic distances, the power state corresponding to the target characteristic distance that meets the set requirements is used as the power state of the target power system.

3. The method according to claim 1, characterized in that The method of determining the centroids of each target cluster as the centroids of each cluster maintained in the basic power state classification network, and performing multiple transfer learning debugging on the basic power state classification network based on the power data learning sample set and the target control network to obtain the debugged target power state classification network includes: Determining each target cluster centroid as each cluster centroid maintained in the basic power state classification network, and repeatedly debugging the basic power state classification network based on the power data learning sample set to obtain a debugged process power state classification network; The process power state classification network is subjected to multiple transfer learning debugging through the power data learning sample set and the target control network to obtain a debugged target power state classification network.

4. The method according to claim 3, characterized in that The basic power state classification network includes a basic characterization vector extraction module and a basic decision module. When the basic power state classification network is repeatedly debugged, it includes: Loading the acquired second power data learning sample into the basic representation vector extraction module to obtain a first basic power state representation vector corresponding to the second power data learning sample; Loading the first basic power state representation vector into the basic decision module to obtain a first basic decision result obtained by reasoning the first basic power state representation vector and the centroid vectors of each target cluster; The network learnable variables of the basic power state classification network are optimized by using the error between the first basic decision result and the power state training indication information corresponding to the second power data learning sample.

5. The method according to claim 4, characterized in that The first basic decision result obtained by reasoning the first basic power state characterization vector and the centroid vectors of each target cluster includes: Determining, by means of the basic decision module, first basic spatial distances between the first basic power state representation vector and the centroids of each target cluster; A first basic decision result estimated for the second power data learning example is determined through each first basic spatial distance.

6. The method according to claim 3, characterized in that The process power state classification network includes a process representation vector extraction module and a process decision module. When a round of transfer learning debugging is performed on the process power state classification network, it includes: Loading the acquired third power data learning sample into the target control network to obtain a first reference power state representation vector extracted by the target control network; Loading the third power data learning sample into the debugged process representation vector extraction module to obtain a corresponding process power state representation vector, and obtaining an estimated process decision result based on the process decision module through the process power state representation vector and the centroids of each target cluster; The network learnable variables of the process power state classification network are optimized by the error between the first reference power state characterization vector and the process power state characterization vector, and by the error between the process decision result and the power state training indication information corresponding to the third power data learning sample.

7. The method according to claim 6, characterized in that The optimizing the network learnable variables of the process power state classification network by the error between the first reference power state characterization vector and the process power state characterization vector, and by the error between the process decision result and the power state training indication information corresponding to the third power data learning example, comprises: Calculating a first transfer learning error result by an error between the first reference power state characterization vector and the process power state characterization vector, and calculating a first classification error result by an error between the process decision result and the power state training indication information corresponding to the third power data learning example; Network learnable variables of the process power state classification network are optimized through the first transfer learning error result and the first classification error result.

8. The method according to claim 1, characterized in that When performing a round of transfer learning debugging on the basic power state classification network, it includes: Loading the acquired fourth power data learning sample into the target control network to obtain a second reference power state representation vector extracted by the target control network, and loading the fourth power data learning sample into the basic power state classification network to obtain an extracted second basic power state representation vector and an estimated second basic decision result; The network learnable variables of the basic power state classification network are optimized by the error between the second reference power state characterization vector and the second basic power state characterization vector, and by the error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning sample.

9. The method according to claim 8, characterized in that The basic power state classification network includes a basic characterization vector extraction module and a basic decision module, and the fourth power data learning sample is loaded into the basic power state classification network to obtain the extracted second basic power state characterization vector and the estimated second basic decision result, including: Loading the fourth power data learning sample into the basic representation vector extraction module to obtain an extracted second basic power state representation vector; Obtaining an estimated second basic decision result through the basic decision module, through the second basic power state characterization vector and the centroids of each target cluster; The optimizing the network learnable variables of the basic power state classification network by the error between the second reference power state characterization vector and the second basic power state characterization vector, and by the error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning example, comprises: Calculate a second transfer learning error result based on an error between the second reference power state characterization vector and the second basic power state characterization vector, and calculate a second classification error result based on an error between the second basic decision result and the power state training indication information corresponding to the fourth power data learning example; The network learnable variables of the basic power state classification network are optimized through the second transfer learning error result and the second classification error result.

10. A power forecasting system, characterized in that: include: processor; and a memory for storing executable instructions for the processor; The processor is configured to perform the method of any one of claims 1 to 9 by executing the executable instructions.

Images