Index system construction method and system for electricity market operation

Abstract

The invention provides an index system construction method and system for electricity market operation, and the method comprises the steps: integrating data from different sources of an electricity market, and carrying out the preprocessing of the data; pre-processed data are obtained, and self-supervised data extraction is carried out on the data from the pre-processed data; constructing a power market multi-task meta-learning optimization model; a real-time operation index system of the electricity market is constructed and dynamically adjusted; a multi-target strategy optimization algorithm is constructed, and an optimal market strategy is designed according to the market indexes adjusted in real time; and carrying out system learning and system adaptive updating according to the optimal market strategy, and keeping the optimal performance. By combining meta-learning and self-supervised learning technologies, the system can quickly adapt to different market environments and scenes, automatically mine and generate key operation indexes from large-scale unmarked power market data, and realize real-time updating and optimization of the index system.

Description

Technical Field

[0001] The present invention belongs to the technical field of power market operation, and in particular relates to a method and system for constructing an indicator system for power market operation. Background Art

[0002] In the current electricity market, the construction and optimization of operating indicators are the key to ensuring the efficient operation of the market. As a complex dynamic system, the electricity market is affected by a variety of uncertain factors (such as load demand, power generation costs, electricity price fluctuations, weather conditions, policy changes, etc.). Traditional methods for constructing electricity market indicator systems mainly rely on expert experience and statistical analysis of historical data. These methods usually use linear models or time series analysis to predict future electricity demand and price fluctuations, and set corresponding operating indicators based on these prediction results. However, these traditional methods have several significant defects.

[0003] First, traditional methods often lack flexibility and real-time performance. Due to the complexity and uncertainty of the power market environment, market conditions may change dramatically in a short period of time (for example, sudden equipment failures, weather changes, policy adjustments, etc.). In these cases, traditional indicator systems based on static rules or fixed models cannot respond quickly to changes, which may lead to inefficient market operation or insufficient power supply, or even cause large-scale power outages and market fluctuations. Second, traditional methods rely on a large amount of historical data and expert knowledge, which not only increases the cost and complexity of building an indicator system, but also these data and experience may no longer be applicable when facing new market conditions, resulting in the obsolescence or failure of the indicator system. In addition, the data types in the power market are diverse and unstructured, including but not limited to time series data (such as load, price), text data (such as policy documents), spatial data (such as geographic location), etc. Traditional methods are difficult to effectively integrate and analyze these multi-source heterogeneous data, and thus cannot fully utilize the potential information of the data to optimize market operations.

[0004] Modern machine learning techniques, such as supervised learning, perform well in some data-rich scenarios, but their generalization and adaptability are limited in environments such as the power market where changes are frequent and data distribution is time-varying. In addition, supervised learning relies on a large amount of labeled data, which severely limits its application in a field such as the power market where data annotation costs are high and labels are difficult to define. Summary of the invention

[0005] The purpose of the present invention is to design a method and system for constructing an indicator system for power market operation, and a method and system for adaptively constructing power market operation indicators based on meta-learning and self-supervised learning to overcome the above-mentioned defects and deficiencies of traditional methods. By combining meta-learning and self-supervised learning technologies, the system can quickly adapt to different market environments and scenarios, and automatically mine and generate key operating indicators from large-scale unlabeled power market data, so as to achieve real-time update and optimization of the indicator system.

[0006] In order to achieve the above object, a first aspect of the present invention provides a method for constructing an indicator system for power market operation, the method comprising the following steps:

[0007] Step 1: Integrate data from different sources in the power market and pre-process the data;

[0008] Step 2: Obtain the preprocessed data and perform self-supervised data extraction from the preprocessed data; specifically, it includes:

[0009] S201, constructing a feature matrix and a target matrix according to the preprocessed data set, wherein the feature matrix is ​​extracted from the preprocessed data set, and the target matrix is ​​generated by a sliding window technology, and is used to predict the market status in the future t time steps;

[0010] S202, constructing a self-supervised loss function according to the error of future state prediction, and then constructing a self-supervised model based on a multi-layer bidirectional long short-term memory network according to the self-supervised loss function, and using a gradient descent algorithm to optimize the model parameters of the self-supervised model;

[0011] S203, using the trained self-supervisory model to generate a new corresponding high-dimensional feature representation;

[0012] Step 3: Build a multi-task meta-learning optimization model for the power market, and use high-dimensional feature representation to train the multi-task meta-learning optimization model for the power market, so that the multi-task meta-learning optimization model for the power market can adapt to the dynamic changes of the power market and optimize various market indicators; specifically:

[0013] S301, construct multiple tasks according to different scenarios in the power market, and use different task-specific sub-feature sets from the high-dimensional feature representation as task data;

[0014] S302, combining task-specific loss terms and innovative inter-task coordination regularization terms to construct a multi-objective loss function of a multi-task meta-learning optimization model, ensuring that the multi-task meta-learning optimization model can adapt to and optimize a variety of market indicators in different task environments; wherein the multi-objective loss function of the multi-task meta-learning optimization model It is expressed as follows:

[0015]

[0016] Where N represents the total number of tasks; Represents task T i The specific loss, W i It is task T i Parameters; α i represents the weight of the task loss, which is used to balance the impact of different tasks; λ represents the task-specific regularization weight; Represents task i for feature H ij The loss gradient is used to evaluate the importance of features; ρ j represents the importance weight of feature j; β represents the weight of the inter-task coordination regularization term; γ jk represents the similarity weight between task j and task k, which is used to control the coordination of task parameters;

[0017] S303, jointly training the multi-task meta-learning optimization model through multiple tasks, so that meta-parameters are shared on different tasks, thereby adapting to new tasks;

[0018] S304, verifying and optimizing the trained multi-task meta-learning optimization model;

[0019] Step 4: Build and dynamically adjust the real-time operation indicator system of the power market; specifically include:

[0020] S401, constructing a feature mapping function according to meta parameters and real-time data streams to generate real-time market indicators;

[0021] S402, introducing an environment change response mechanism into the real-time data stream to detect changes in the real-time data stream and adaptively adjust indicators;

[0022] S403, designing an adaptive feedback optimization mechanism to update the index mapping matrix and the parameters of the feature mapping function in the index multi-task meta-learning optimization model, and finally constructing a total loss function;

[0023] Step 5: Construct a multi-objective strategy optimization algorithm to design the optimal market strategy based on the real-time adjusted market indicators; specifically, it includes:

[0024] According to the market status and the initial strategy set at the time of historical data generation, a multi-objective optimization function is constructed based on the initial strategy set combined with the weighted sum of multiple market objectives, and the strategy set is iteratively updated using the optimization algorithm according to the multi-objective optimization function to obtain the optimal market strategy, and a strategy evaluation function is constructed to evaluate the optimal market strategy, and the weights in the optimization objective function are dynamically adjusted according to the evaluation results to dynamically update the market strategy;

[0025] Step 6: Perform system learning and adaptive updates based on the optimal market strategy to maintain optimal performance.

[0026] Further, the preprocessing includes aligning the time series using a dynamic time correction method;

[0027] The Step 1 further includes:

[0028] For time-sequence data, a sliding window is used to calculate local features; for unstructured data, an embedded semantic transformation method is used to extract useful information;

[0029] A dynamic missing value filling method based on multivariate interpolation is used to fill missing values. At the same time, an anomaly detection method based on adaptive threshold is used to dynamically set the detection threshold according to the local statistical characteristics of the data. If the data is greater than the detection threshold, the current data is marked as an outlier.

[0030] Furthermore, in the self-supervised loss function The time difference constraint is also introduced in the self-supervisory loss function, and the self-supervised loss function is expressed as follows:

[0031]

[0032] in, Represents the target matrix predicted by the model, G i represents the actual target matrix, λ represents the weight of the time difference constraint, It is used to measure the prediction changes between consecutive time steps, ensuring that the model not only learns the point-to-point prediction ability of the data, but also captures the smoothness and consistency in time; n represents the number of samples; t represents the time step; t' represents the total time step;

[0033] The self-supervisory model based on a multi-layer bidirectional long short-term memory network is constructed according to the self-supervisory loss function, specifically including:

[0034] Randomly initialize the weight parameters of the self-supervised model based on a multi-layer bidirectional long short-term memory network;

[0035] The forward propagation function of the multi-layer bidirectional long short-term memory network is used to calculate the predicted output of each sample;

[0036] Calculate the gradient according to the self-supervised loss function, and then use the back-propagation algorithm to update the weight parameters of the self-supervised model;

[0037] Through multiple rounds of iterations, minimize the self-supervised loss function Until convergence.

[0038] Furthermore, in S301, a task diversity factor is added when constructing multiple tasks to ensure that the input features of multiple task subsets remain different;

[0039] The S303 specifically includes:

[0040] Randomly initialize the meta-parameters of the multi-task meta-learning optimization model and set the initial learning rate and meta-learning rate;

[0041] For each meta-learning iteration, randomly select a task from the task set, randomly sample a mini-batch of data from the task-specific subset of inputs, calculate the task loss, and update the task-specific parameters via gradient descent;

[0042] After all tasks are updated, the meta-loss of all tasks is calculated and the meta-parameters are updated by gradient descent.

[0043] Repeat the above steps until the meta-parameters converge or meet the preset optimization conditions.

[0044] Furthermore, the S304 specifically includes:

[0045] Using a set of validation tasks Apply the trained meta-parameters W meta , calculate the validation loss for each task

[0046] Weight the task loss according to the validation loss α i , feature importance weight ρ j and task similarity weight γ jk Perform dynamic adjustments and train again to optimize model performance, as shown below:

[0047]

[0048] in, represents the adjusted task loss weight; represents the adjusted feature importance weight; represents the adjusted task similarity weight; corr(H j ,H k ) represents the feature set H j and H k The correlation coefficient between .

[0049] Furthermore, the S401 specifically includes:

[0050] Design a feature mapping function to convert the real-time data stream into a high-dimensional feature space, and use the indicator mapping matrix extracted from the multi-task meta-learning optimization model to map the feature vector mapped to the high-dimensional space to the indicator space to generate the initial market indicator vector;

[0051] Among them, the environmental change response mechanism senses the intensity of changes in the market environment by calculating the change rate of real-time data. If the change rate of real-time data exceeds a preset threshold, it is considered that the market environment has changed significantly and the indicator system needs to be adjusted.

[0052] Furthermore, the total loss function Including indicator adjustment loss function And the regularization term Ω(t), expressed as follows:

[0053]

[0054] in, represents the stability control term, which is used to prevent instability caused by frequent adjustments, and γ represents the weight factor of stability control;

[0055] The indicator adjusts the loss function Contains real-time indicator errors, expressed as follows:

[0056]

[0057] Among them, M i (t) represents the i-th market indicator currently calculated; represents the target i-th market indicator, which is derived based on historical data and forecasting models; λ1 represents the time smoothing regularization weight; m represents the number of features; Δt represents the time change.

[0058] Furthermore, the multi-objective optimization function It is expressed as follows:

[0059]

[0060] Where n represents the dimension of the strategy set; g k (U(t),S(t)) represents the kth objective function associated with strategy U(t) and market state S(t); w k is the weight of the kth objective function, which is used to balance the importance of different objectives.

[0061] Furthermore, the system learning and system adaptive updating according to the optimal market strategy specifically include:

[0062] S601, construct an adaptive memory module to store past strategy data and market status information, the adaptive memory module includes a dynamically updated memory matrix M mem (t)∈R m×k , used to save the market state vector S(t) and the corresponding strategy U(t) of the most recent k times. The update of the memory matrix is ​​based on a rolling window mechanism. Each time it is updated, the earliest set of data is removed and the latest data is added.

[0063] S602: Use the current market strategy U(t+1) and the historical information in the adaptive memory module to construct a new adaptive loss function At the same time, the difference between the current market status and historical data is combined to minimize the error between the prediction of the multi-task meta-learning optimization model and the actual market; the adaptive loss function It is expressed as follows:

[0064]

[0065] in, is the j-th market state value predicted by the model; S j (t+1) is the actual j-th market state value; γ is a weight factor that controls the impact of past market state differences; i is the index in the past k times, which is used to calculate the weight of past data;

[0066] S603: Based on the strategy evaluation results and the historical strategy data of the adaptive memory module, a strategy update function φ(·) is constructed to adjust the strategy in real time according to market changes, as shown below:

[0067]

[0068] Among them, U new (t+1) represents the updated strategy vector; β represents the strategy adjustment step size, which is used to control the strength of the strategy update; Represents the gradient of the policy evaluation function, which is used to adjust the policy;

[0069] S604. Repeatedly adjust and optimize strategies and system parameters to maintain the adaptability and robustness of the system.

[0070] In a second aspect of the present invention, a system for constructing an indicator system for power market operation is provided, the system comprising the following units:

[0071] The power market data collection unit is used to integrate data from different sources in the power market and pre-process the data;

[0072] The data self-supervision unit is used to obtain the pre-processed data and perform self-supervision data extraction from the pre-processed data; specifically, it includes:

[0073] S201, constructing a feature matrix and a target matrix according to the preprocessed data set, wherein the feature matrix is ​​extracted from the preprocessed data set, and the target matrix is ​​generated by a sliding window technology, and is used to predict the market status in the future t time steps;

[0074] S202, constructing a self-supervised loss function according to the error of future state prediction, and then constructing a self-supervised model based on a multi-layer bidirectional long short-term memory network according to the self-supervised loss function, and using a gradient descent algorithm to optimize the model parameters of the self-supervised model;

[0075] S203, using the trained self-supervisory model to generate a new corresponding high-dimensional feature representation;

[0076] The data meta-learning unit is used to build a multi-task meta-learning optimization model for the power market and use high-dimensional feature representation to train the multi-task meta-learning optimization model for the power market, so that the multi-task meta-learning optimization model for the power market can adapt to the dynamic changes of the power market and optimize multiple market indicators; specifically, it includes:

[0077] S301, construct multiple tasks according to different scenarios in the power market, and use different task-specific sub-feature sets from the high-dimensional feature representation as task data;

[0078] S302, combining task-specific loss terms and innovative inter-task coordination regularization terms to construct a multi-objective loss function of a multi-task meta-learning optimization model, ensuring that the multi-task meta-learning optimization model can adapt to and optimize a variety of market indicators in different task environments; wherein the multi-objective loss function of the multi-task meta-learning optimization model It is expressed as follows:

[0079]

[0080] Where N represents the total number of tasks; Represents task T i The specific loss, W i It is task T i Parameters; α i represents the weight of the task loss, which is used to balance the impact of different tasks; λ represents the task-specific regularization weight; Represents task i for feature H ij The loss gradient is used to evaluate the importance of features; ρ j represents the importance weight of feature j; β represents the weight of the inter-task coordination regularization term; γ jk represents the similarity weight between task j and task k, which is used to control the coordination of task parameters;

[0081] S303, jointly training the multi-task meta-learning optimization model through multiple tasks, so that meta-parameters are shared on different tasks, thereby adapting to new tasks;

[0082] S304, verifying and optimizing the trained multi-task meta-learning optimization model;

[0083] The data indicator system construction unit is used to construct and dynamically adjust the real-time operation indicator system of the power market; specifically includes:

[0084] S401, constructing a feature mapping function according to meta parameters and real-time data streams to generate real-time market indicators;

[0085] S402, introducing an environment change response mechanism into the real-time data stream to detect changes in the real-time data stream and adaptively adjust indicators;

[0086] S403, designing an adaptive feedback optimization mechanism to update the index mapping matrix and the parameters of the feature mapping function in the index multi-task meta-learning optimization model, and finally constructing a total loss function;

[0087] The indicator system optimization unit is used to construct a multi-objective strategy optimization algorithm and design the optimal market strategy based on the market indicators adjusted in real time; specifically, it includes:

[0088] According to the market status and the initial strategy set at the time of historical data generation, a multi-objective optimization function is constructed based on the initial strategy set combined with the weighted sum of multiple market objectives, and the strategy set is iteratively updated using the optimization algorithm according to the multi-objective optimization function to obtain the optimal market strategy, and a strategy evaluation function is constructed to evaluate the optimal market strategy, and the weights in the optimization objective function are dynamically adjusted according to the evaluation results to dynamically update the market strategy;

[0089] The system optimization unit is used to perform system learning and system adaptive updates based on the optimal market strategy to maintain optimal performance.

[0090] The beneficial technical effects of the present invention are at least as follows:

[0091] The present invention proposes a method and system for adaptively constructing power market operation indicators based on meta-learning and self-supervised learning to overcome the above-mentioned defects and shortcomings of traditional methods. By combining meta-learning and self-supervised learning technologies, the system can quickly adapt to different market environments and scenarios, and automatically mine and generate key operation indicators from large-scale unlabeled power market data, realizing real-time updating and optimization of the indicator system.

[0092] Specifically, the introduction of the meta-learning module enables the system to quickly adapt to market changes. Meta-learning learns how to adjust and optimize the market indicator system through training in multiple tasks and environments. When the market environment changes (such as policy adjustments, emergencies, etc.), the system can quickly retrain with a small amount of new market data to build a new optimal indicator system, thereby improving the flexibility and real-time performance of the system. This innovation solves the problem of lack of flexibility and real-time performance of traditional methods, allowing power market operations to adapt to market changes more dynamically and accurately, avoiding market instability and inefficiency caused by slow response.

[0093] On the other hand, the application of the self-supervised learning module enables the system to automatically extract meaningful features and indicators from large-scale unlabeled electricity market data without a large amount of labeled data. In this way, the system is able to automatically generate a series of key operating indicators, such as the cost-effectiveness ratio of power generation, real-time load forecasting error, anomaly detection frequency, etc. This ability to automate feature extraction and indicator generation significantly reduces the reliance on expert knowledge and manual labeling, reduces the cost and complexity of building an indicator system, and solves the problem that traditional methods are difficult to integrate and analyze multi-source heterogeneous data. In addition, self-supervised learning can continuously learn and adjust, so that the indicator system can be updated in real time according to market dynamics (such as demand changes, supply fluctuations), ensuring its foresight and accuracy.

[0094] In summary, the present invention provides an innovative method for constructing an electricity market operation indicator system through the integrated application of meta-learning and self-supervised learning. It not only solves the defects of traditional methods, but also introduces automated and intelligent feature extraction and indicator generation capabilities, greatly improving the operational efficiency and stability of the electricity market. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0096] Figure 1 The present invention is a flowchart of a method for constructing an indicator system for power market operation according to an embodiment of the present invention.

[0097] Figure 2 The present invention is a system framework diagram for constructing an indicator system for power market operation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0098] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0099] In one or more embodiments, Figure 1 As shown, the present invention discloses a method for constructing an indicator system for power market operation, the method comprising steps 1 to 6, including:

[0100] Step 1: Integrate data from different sources in the electricity market and pre-process the data.

[0101] Specifically, in order to adapt to the complexity and diverse data characteristics of the power market, in the data collection and preprocessing steps, the present invention needs to design a preprocessing solution for the specific needs of the power market. The data sources of the power market are extensive, including both time series data (such as load forecasts, electricity price fluctuations) and unstructured data (such as policy documents, weather forecasts).

[0102] First, integrate data from different sources, such as historical transaction data of the electricity market (electricity price P t , Power Q t ), weather data (temperature T t 、Humidity t ), policy documents (in text form), etc. Since the timestamps of these data are inconsistent, time alignment is required. The present invention uses a dynamic time correction method (Dynamic Time Correction, DTC) to align time series data:

[0103]

[0104] in, represents the aligned data points, X i,t is the original data point, Δt is the global standard time interval, Δt i is the time interval of each data source. Through dynamic time correction, data from different time scales can be effectively aligned to ensure the consistency of subsequent analysis.

[0105] Furthermore, according to the characteristics of the power market, the present invention designs a set of feature extraction methods. Specifically, for time series data (such as electricity price and load), the present invention uses a sliding window to calculate local features, such as the average value μ X (t), standard deviation σ X (t) and skewness γ X (t) etc.:

[0106]

[0107] Among them, X t represents the data value at time t, and N is the window size. These local statistical features can capture the short-term dynamic changes of power market data and provide valuable input for self-supervised learning models.

[0108] Furthermore, for unstructured data (such as policy documents and weather forecasts), the present invention uses the Embedded Semantic Transformation (EST) method to extract effective information from it. The specific steps are as follows:

[0109] First, convert the text data into a vector representation v p A custom embedding generation model is used that combines contextual relevance and word vectors (WordVectors) of electricity market-specific vocabulary to generate v p :

[0110]

[0111] Among them, w j Represents the word vector of the jth word in the text, and M is the total number of words in the text.

[0112] Then, a topic-based clustering method is applied to cluster these embedding vectors to form high-level semantic features:

[0113]

[0114] Among them, C k is the kth cluster, is the center of the cluster.

[0115] Among them, the embedding generation and topic clustering steps can transform unstructured data into actionable features and provide rich contextual information for subsequent analysis.

[0116] Furthermore, there may be missing values ​​and outliers in the power market data. To solve this problem, the present invention adopts a dynamic missing value filling method based on multivariate interpolation (Dynamic Multivariate Interpolation, DMI) and an anomaly detection method based on adaptive threshold (Adaptive Threshold Anomaly Detection, ATAD):

[0117] Missing value filling: Use the nearest neighbor interpolation method and the local weighted regression method to fill the missing data:

[0118]

[0119] Among them, X k,t is the value of the K points closest to the missing point, w k is the corresponding weighting coefficient.

[0120] Anomaly detection: Dynamically set the detection threshold θ(t) based on the local statistical characteristics of the data:

[0121] θ(t)=μ X (t)+3σ X (t)

[0122] If X t >θ(t), then X t These steps ensure the integrity and accuracy of the data and reduce the impact of noise on model training.

[0123] Furthermore, the dimensionality of the power market data may be very high and contain a lot of redundant information. In order to reduce the computational complexity and improve the efficiency of model training, the present invention adopts a method combining principal component analysis (PCA) and maximum information coefficient (MIC) to perform dimensionality reduction and feature selection:

[0124] Principal component analysis: Calculate the eigenvalue decomposition of the data covariance matrix and select the eigenvectors corresponding to the first d largest eigenvalues ​​as the principal components:

[0125] Y=XW

[0126] Among them, X is the preprocessed data matrix, W is the feature vector matrix, and Y is the data matrix after dimensionality reduction.

[0127] Maximum information coefficient: On the reduced-dimensional data, the maximum information coefficient between each feature and the target variable is calculated, and those features with the strongest relationship with the target variable are selected as the final input features.

[0128] Among them, the dimensionality reduction and feature selection steps can remove redundant information, reduce the dimension of the data, retain the features that have the greatest impact on the operation of the power market, and improve the computational efficiency and predictive ability of the model.

[0129] Step 2: Obtain the preprocessed data and perform self-supervised data extraction from the preprocessed data.

[0130] Specifically, the preprocessed data set X obtained in step 1 is assumed to have a dimension of (n, m), where n represents the number of samples and m represents the number of features. The present invention first constructs an input feature matrix and the target matrix Here k is the dimension of the input features and l is the dimension of the target matrix. The input feature matrix F is constructed by selecting a set of key features from X (such as electricity price, electricity consumption, weather data, etc.), while the target matrix G is generated by a sliding window technique, aiming to predict the market state in the future t time steps.

[0131] F=[X :,1:k ], G = SlidingWindow (X :,k+1:m ,t)

[0132] Among them, X :,1:k represents the first k columns selected from the original dataset X as input features, X :,k+1:m The remaining columns are used to generate the target matrix. The function SlidingWindow(·, t) is a sliding window operation used to create the target matrix G, which is used for future t time step predictions.

[0133] Furthermore, in order to enable the model to effectively learn meaningful features from the data, the present invention defines a self-supervised loss function. The loss function is based on the error of future state prediction and contains a novel time-dependent regularization term called Temporal Difference Constraint (TDC). Loss Function is defined as follows:

[0134]

[0135] in, is the target matrix predicted by the model, G i is the actual target matrix, λ is the weight of the time difference constraint, It is used to measure the change in predictions between consecutive time steps, ensuring that the model not only learns the point-to-point prediction ability of the data, but also captures the smoothness and consistency over time.

[0136] Furthermore, based on the defined self-supervised loss function, the present invention uses a gradient descent algorithm to optimize the model parameters. The model structure uses a multi-layer bidirectional long short-term memory network (Bi-DirectionalLSTM, Bi-LSTM) to capture the forward and backward dependencies of time series data. The training process of the model is as follows:

[0137] Initialize model parameters: Randomly initialize the weight parameter W of the Bi-LSTM model lstm .

[0138] Forward propagation: Calculate the predicted output for each sample Here Bi-LSTM(·) is the forward propagation function of the Bi-LSTM model.

[0139] Loss calculation and back propagation: According to the self-supervised loss function Computing Gradients Then use the back-propagation algorithm to update the model parameters W lstm :

[0140]

[0141] Where η is the learning rate. Through multiple rounds of iterations, the loss function is continuously minimized Until convergence.

[0142] Furthermore, once the model training is completed, the trained Bi-LSTM model is used to generate new feature representations. Given a new input data set X', the present invention calculates its high-dimensional feature representation H through the Bi-LSTM model:

[0143] H = Bi-LSTM(X';W lstm )

[0144] in, n' is the number of samples in the new dataset, and d is the dimension of the feature representation. This new feature representation H will be used for subsequent meta-learning model training.

[0145] Step 3. Construct a multi-task meta-learning optimization model for the electricity market, and use high-dimensional feature representation to train the multi-task meta-learning optimization model for the electricity market, so that the multi-task meta-learning optimization model for the electricity market can adapt to the dynamic changes of the electricity market and optimize various market indicators.

[0146] Specifically, based on the specific needs and characteristics of the power market, the present invention constructs multiple task environments to train the meta-learning model. i Represents a typical scenario in the power market (e.g., peak load, off-peak load, emergencies, policy changes, etc.). In order to generate different tasks, the present invention extracts different task-specific subsets from the self-supervised learning feature representation H As a task T i The present invention assumes that the dimension of the self-supervised feature H is (n, d), where n is the number of samples and d is the number of features. In order to capture the diversity of different market scenarios, the present invention introduces the task diversity factor δ in the task division process. i , ensuring that the input features of each task subset remain distinct:

[0147] H i =TaskSample(H,δ i )

[0148] Among them, TaskSample(H,δ i ) is a task sampling function, which sets the task diversity factor δ i Different feature subsets are randomly extracted from H to construct the task environment.

[0149] Furthermore, the meta-learning model aims to optimize multiple task objectives (such as electricity price prediction accuracy, load management, and market stability) at the same time. To this end, the present invention introduces a multi-objective loss function It combines task-specific loss terms and innovative inter-task coordination regularization (ITCR) to ensure that the model can quickly adapt and optimize multiple market indicators in different task environments. The definition is as follows:

[0150]

[0151] Where N is the total number of tasks; Represents task T i The specific loss, W i It is task T i Parameters; α i is the weight of the task loss, which is used to balance the impact of different tasks; λ is the task-specific regularization weight; is the task i for feature H ij The loss gradient is used to evaluate the importance of features; ρ j is the importance weight of feature j; β is the weight of the inter-task coordination regularization term; γ jk represents the similarity weight between task j and task k, which is used to control the coordination of task parameters.

[0152] The inter-task coordination regularization term (ITCR) and task-specific gradient regularization term introduced in this formula ensure that the model not only optimizes the performance of a single task, but also coordinates the relationship between multiple tasks, ensuring the optimal balance between parameter sharing and feature utilization between tasks.

[0153] Furthermore, the meta-learning model optimizes the joint training of multiple tasks so that the model parameters are shared across different tasks, thereby quickly adapting to new tasks. The training steps are as follows:

[0154] Initialize meta parameters: Randomly initialize the model’s meta parameters W meta . Set the initial learning rate η and meta learning rate η'.

[0155] Task sampling and training: For each meta-learning iteration, from the task set {T i} Randomly select a task T i, from the task-specific input H i Randomly sample a mini-batch of data from the , and calculate the loss And update the task-specific parameters W by gradient descent i :

[0156]

[0157] Meta parameter update: After all tasks are updated, calculate the meta loss of all tasks And update the meta-parameter W by gradient descent meta :

[0158]

[0159] Iterative optimization: Repeat the above steps until the meta-parameter W meta Converge or meet the preset optimization conditions.

[0160] Furthermore, after training is completed, the meta-learning model is verified and optimized to ensure its effectiveness in various market scenarios. Apply the trained model parameters W meta , calculate the validation loss for each task Weight the task loss according to the validation loss α i , feature importance weight ρ j and task similarity weight γ jk Perform dynamic adjustments and train again to further optimize model performance:

[0161]

[0162] in, is the adjusted task loss weight; is the adjusted feature importance weight; is the adjusted task similarity weight; corr(H j ,H k ) is the feature set H j and H k The correlation coefficient between .

[0163] The meta-learning model can quickly adapt to and optimize the operating indicators of the power market in different task environments. The present invention first constructs a multi-task learning environment and uses the task diversity factor δ i Generate different task inputs. Then, by defining a multi-objective loss function The combination of task-specific gradient regularization terms and inter-task coordination regularization terms ensures that the model performs well when optimizing multiple market indicators. During the training process, the model learns the common features of multiple tasks through steps such as task training and updating, and meta-parameter updating. Finally, the generalization ability and adaptability of the model are further improved through the verification and adaptive optimization steps. These features and optimization processes enable the meta-learning model to maintain efficient operation in the rapidly changing electricity market and provide a reliable foundation for subsequent steps.

[0164] Step 4: Build and dynamically adjust the real-time operation indicator system of the electricity market.

[0165] Specifically, use the meta-learning model parameters W obtained from step 3 meta The feature mapping function φ(·) is constructed with the real-time data stream D(t) to generate real-time market indicators. Considering the nonlinear and time-varying characteristics of power market data, the present invention designs a feature mapping function to convert the data into a high-dimensional feature space:

[0166] F(t)=φ(D(t),W meta )=σ(W2·σ(W1·D(t)+b1)+b2)

[0167] in, is the feature vector mapped to the high-dimensional space. is the real-time data input. W1, W2 are the weight matrices obtained from the meta-learning model. b1, b2 are bias terms. σ(·) is the activation function (such as ReLU or Sigmoid).

[0168] Then, the feature vector F(t) is mapped to the indicator space using the initial indicator mapping matrix A extracted from the meta-learning model to generate the initial market indicator vector

[0169] M(t)=A·F(t)+c

[0170] in, is the initial market indicator vector. A∈R m×q is the indicator mapping matrix derived from the meta-learning model. c∈R m is the bias term.

[0171] Furthermore, in order to cope with the dynamic changes in the power market, an environmental change response mechanism (ECRM) is introduced to detect changes in real-time data and adaptively adjust indicators. Change detection senses the intensity of changes in the market environment by calculating the rate of change of real-time data:

[0172]

[0173] Where ξ(t) is the dynamic perception coefficient, which is used to measure the intensity of changes in the market environment. D(t) and D(t-Δt) are the real-time data inputs at the current and previous moments, respectively. Δt is the time interval. If ξ(t) exceeds the preset threshold θ ξ , it is considered that the market environment has changed significantly and the indicator system needs to be adjusted.

[0174] Furthermore, in the process of index adjustment, the present invention designs a core optimization formula, which combines the real-time error and the regularization term to ensure the robustness of the adjustment. This core formula is used to update the parameters of the index mapping matrix A and the feature mapping function φ(·):

[0175]

[0176] Where A'(t) is the updated indicator mapping matrix. η is the learning rate, which controls the update step size. It is about the indicator adjustment loss function The gradient of . λR(A) is a regularization term used to prevent over-adjustment. R(A) can be a sparse regularization (such as L1 regularization) or a time smoothing regularization (such as difference regularization). Indicator Adjustment Loss Function Contains real-time indicator errors and regularization terms:

[0177]

[0178] Among them, M i (t0 is the i-th market indicator currently calculated. is the target i-th market indicator, based on historical data and forecasting models. λ1 is the time smoothing regularization weight. By minimizing The system can dynamically adjust the indicator mapping matrix A to ensure that the market indicator system remains stable and efficient in a changing environment.

[0179] Finally, the comprehensive adjusted loss And regularization term, forming the overall optimization objective function

[0180]

[0181] in: is a stability control term used to prevent instability caused by frequent adjustments. γ is a weight factor for stability control. By minimizing The system maintains a certain stability while responding quickly to market changes.

[0182] Step 5: Build a multi-objective strategy optimization algorithm and design the optimal market strategy based on real-time adjusted market indicators.

[0183] Specifically, the current market state is modeled based on the real-time market indicator M(t) obtained from step S4. Define the market state vector S(t)∈R m , which is composed of the real-time market indicator vector M(t). Then, the market strategy is initialized using the market state vector S(t) and historical market data Where U(t) represents the set of strategies at time t (such as power generation scheduling strategy, price setting strategy, etc.). The initial strategy setting is achieved through a strategy generation function ψ(·), which generates the initial strategy based on market status and historical data:

[0184] U(t)=ψ(S(t),H)

[0185] Where H represents the historical market data set, ψ(·) is the strategy generation function, which generates an initial strategy that adapts to the current market environment by combining historical data and the current market status.

[0186] Furthermore, the goal of power market strategy optimization is to maximize market efficiency (such as economic benefits), stability (such as price volatility) and sustainability (such as green energy utilization) at the same time. To this end, the present invention introduces a multi-objective optimization function It combines the weighted sum of multiple market objectives. The optimization function is defined as follows:

[0187]

[0188] Where n is the dimension of the policy set. k (U(t),S(t)) is the kth objective function (e.g., economic return, price stability, etc.) associated with the strategy U(t) and the market state S(t). k is the weight of the kth objective function, which is used to balance the importance of different objectives.

[0189] Furthermore, according to the optimization function The strategy is iteratively updated using the gradient descent method or other optimization algorithms. In order to ensure the adaptability and effectiveness of the strategy in different market environments, the present invention designs a strategy feedback optimization mechanism (SFOM). After each strategy update, real-time correction is performed based on the market feedback data F(t):

[0190]

[0191] Among them, U(t+1) is the update policy at time t+1. η is the policy learning rate, which controls the step size of the policy update. is the optimization function The gradient of the policy U. λR(U) is the policy regularization term, which is used to prevent the policy from being over-optimized or deviating from actual market demand.

[0192] Furthermore, through the strategy feedback optimization mechanism, the present invention evaluates the effectiveness of the strategy after each strategy update to ensure that the strategy can meet the actual needs and changes of the current market. To this end, the present invention introduces a strategy evaluation function ε(U(t), S(t)) to evaluate the performance of the strategy under the current market state and dynamically adjust the weight w in the optimization objective function according to the evaluation results. k :

[0193]

[0194] in, is the weight of the kth objective function after adjustment. k (U(t), S(t)) is the strategy evaluation function for the kth objective, which measures the contribution of the strategy to the objective. By dynamically adjusting the objective weights, the strategy optimization algorithm can adapt to the ever-changing market and ensure that the optimized market strategy is always consistent with the actual market situation.

[0195] It can be understood that the application of indicators and market strategy optimization can dynamically formulate and optimize power market strategies based on the real-time market indicator system. First, the initial strategy is generated using real-time market status and historical data. Then, the multi-objective optimization function is used to Integrate multiple market objectives to ensure that the strategy strikes a balance between market efficiency, stability, and sustainability. Then, use the strategy feedback optimization mechanism (SFOM) to update and correct the strategy in real time to ensure that the strategy can adapt to the actual needs of the market. Finally, through strategy evaluation and re-optimization, dynamically adjust the weights of the optimization objectives so that the strategy can be continuously optimized and remain effective. This process realizes the dynamic optimization and adaptive adjustment of the power market strategy, improving the overall performance and operating efficiency of the market.

[0196] Step 6: Perform system learning and adaptive updates based on the optimal market strategy to maintain optimal performance.

[0197] Specifically, the output of step S5, namely the optimized market strategy U(t+1) and the strategy evaluation result To carry out continuous learning and system adaptive update. The purpose of continuous learning is to allow the system to automatically adjust its own model and strategy in a changing market environment to maintain optimal performance. The present invention designs an adaptive multi-task continuous learning algorithm (Adaptive Multi-Task Continuous Learning Algorithm, AMT-CLA), which can adaptively update model parameters and strategies in real time to ensure the intelligence and flexibility of power market operation.

[0198] Specifically, an adaptive memory module (AMM) is introduced to store past strategy data and market status information. This module contains a dynamically updated memory matrix It is used to save the market state vector S(t) and the corresponding strategy U(t) for the most recent k times. The update of the memory matrix is ​​based on a rolling window mechanism. Each time it is updated, the earliest set of data is removed and the latest data is added. The update formula of the memory matrix is ​​as follows:

[0199] M mem (t)=[M mem (t-1) :,2:k ,[S(t),U(t)]]

[0200] Among them, M mem (t-1) :,2:k represents the memory matrix after removing the first column, and [S(t), U(t)] represents the current market status and strategy.

[0201] Furthermore, in order to enable the system model to continuously adapt to market changes, the present invention adopts an adaptive parameter update method based on meta-learning. Using the current market strategy U(t+1) and the historical information in the memory module, the present invention defines a new adaptive loss function This loss function combines the difference between the current market state and historical data, aiming to minimize the error between the model prediction and the actual market. The formula of the adaptive loss function is as follows:

[0202]

[0203] in, is the j-th market state value predicted by the model, S j (t+1) is the actual j-th market state value. γ is a weight factor that controls the impact of past market state differences. i is the index in the past k times, which is used to calculate the weight of past data.

[0204] Then, by minimizing the adaptive loss function Update the model parameters W meta :

[0205]

[0206] Among them, η is the learning rate, which is used to control the step size of parameter update.

[0207] Further, using the strategy evaluation result in step S5 Based on the historical strategy data in the memory module, the present invention defines a strategy update function φ(·) to adjust the strategy in real time according to market changes. The core of strategy adjustment is to optimize the flexibility and adaptability of the strategy. The new strategy U new The update formula for (t+1) is as follows:

[0208]

[0209] Among them, U new (t+1) is the updated policy vector. β is the policy adjustment step size, which is used to control the intensity of policy update. is the gradient of the policy evaluation function and is used to adjust the policy.

[0210] Furthermore, to ensure that the system can maintain optimal performance under different market conditions, the present invention introduces a continuous learning mechanism (CLM) to maintain the adaptability and robustness of the system by repeatedly adjusting and optimizing strategies and model parameters. The mechanism is based on a rolling window adaptive update method to update the memory module, model parameters and market strategy in real time, ensuring that the system remains optimized in a changing market environment.

[0211] The final adaptive update objective function Combines adaptive loss and policy optimization objectives:

[0212]

[0213] Among them, λ is the weight coefficient, which controls the contribution of strategy adjustment to the overall goal.

[0214] It can be understood that continuous learning and system adaptive updating realize the continuous optimization and intelligent adaptive adjustment of the power market operation system. First, the adaptive memory module (AMM) records and updates historical data to provide a basis for the adaptive adjustment of models and strategies. Then, the adaptive loss function is used to Dynamically update model parameters so that the model can respond to market changes in a timely manner. Then, the strategy is optimized in real time according to market feedback through the strategy adjustment function φ(·). Finally, the continuous learning mechanism (CLM) ensures that the system maintains the best state in a rapidly changing market environment, realizes the continuous optimization and adaptive adjustment of the power market strategy, and improves the overall performance and operation efficiency of the market.

[0215] In one or more embodiments, Figure 2 As shown, the present invention discloses an indicator system construction system for power market operation, and the system includes the following units:

[0216] The power market data collection unit 101 is used to integrate the data from different sources of the power market and pre-process the data;

[0217] The data self-supervision unit 102 is used to obtain the pre-processed data and perform self-supervision data extraction from the pre-processed data; specifically, it includes:

[0218] S201, constructing a feature matrix and a target matrix according to the preprocessed data set, wherein the feature matrix is ​​extracted from the preprocessed data set, and the target matrix is ​​generated by a sliding window technology, and is used to predict the market status in the future t time steps;

[0219] S202, constructing a self-supervised loss function according to the error of future state prediction, and then constructing a self-supervised model based on a multi-layer bidirectional long short-term memory network according to the self-supervised loss function, and using a gradient descent algorithm to optimize the model parameters of the self-supervised model;

[0220] S203, using the trained self-supervisory model to generate a new corresponding high-dimensional feature representation;

[0221] The data meta-learning unit 103 is used to construct a multi-task meta-learning optimization model for the power market, and use high-dimensional feature representation to train the multi-task meta-learning optimization model for the power market, so that the multi-task meta-learning optimization model for the power market can adapt to the dynamic changes of the power market and optimize multiple market indicators; specifically, it includes:

[0222] S301, construct multiple tasks according to different scenarios in the power market, and use different task-specific sub-feature sets from the high-dimensional feature representation as task data;

[0223] S302, combining task-specific loss terms and innovative inter-task coordination regularization terms to construct a multi-objective loss function of a multi-task meta-learning optimization model, ensuring that the multi-task meta-learning optimization model can adapt to and optimize a variety of market indicators in different task environments; wherein the multi-objective loss function of the multi-task meta-learning optimization model It is expressed as follows:

[0224]

[0225] Where N represents the total number of tasks; Represents task T i The specific loss, W i It is task T iParameters; α i represents the weight of the task loss, which is used to balance the impact of different tasks; λ represents the task-specific regularization weight; Represents task i for feature H ij The loss gradient is used to evaluate the importance of features; ρ j represents the importance weight of feature j; β represents the weight of the inter-task coordination regularization term; γ jk represents the similarity weight between task j and task k, which is used to control the coordination of task parameters;

[0226] S303, jointly training the multi-task meta-learning optimization model through multiple tasks, so that meta-parameters are shared on different tasks, thereby adapting to new tasks;

[0227] S304, verifying and optimizing the trained multi-task meta-learning optimization model;

[0228] The data indicator system construction unit 104 is used to construct and dynamically adjust the real-time operation indicator system of the power market; specifically, it includes:

[0229] S401, constructing a feature mapping function according to meta parameters and real-time data streams to generate real-time market indicators;

[0230] S402, introducing an environment change response mechanism into the real-time data stream to detect changes in the real-time data stream and adaptively adjust indicators;

[0231] S403, designing an adaptive feedback optimization mechanism to update the index mapping matrix and the parameters of the feature mapping function in the index multi-task meta-learning optimization model, and finally constructing a total loss function;

[0232] The indicator system optimization unit 105 is used to construct a multi-objective strategy optimization algorithm and design the optimal market strategy according to the market indicators adjusted in real time; specifically, it includes:

[0233] According to the market status and the initial strategy set at the time of historical data generation, a multi-objective optimization function is constructed based on the initial strategy set combined with the weighted sum of multiple market objectives, and the strategy set is iteratively updated using the optimization algorithm according to the multi-objective optimization function to obtain the optimal market strategy, and a strategy evaluation function is constructed to evaluate the optimal market strategy, and the weights in the optimization objective function are dynamically adjusted according to the evaluation results to dynamically update the market strategy;

[0234] The system optimization unit 106 is used to perform system learning and system adaptive updating according to the optimal market strategy to maintain the best performance.

[0235] These are just some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and making equivalent changes according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for constructing an indicator system for power market operation, characterized in that: The method comprises the following steps: Step 1: Integrate data from different sources in the power market and pre-process the data; Step 2: Obtain the preprocessed data and perform self-supervised data extraction from the preprocessed data; specifically, it includes: S201, constructing a feature matrix and a target matrix according to the preprocessed data set, wherein the feature matrix is ​​extracted from the preprocessed data set, and the target matrix is ​​generated by a sliding window technique, and is used to predict the market status in the future t time steps; S202, constructing a self-supervised loss function according to the error of future state prediction, and then constructing a self-supervised model based on a multi-layer bidirectional long short-term memory network according to the self-supervised loss function, and using a gradient descent algorithm to optimize the model parameters of the self-supervised model; S203, using the trained self-supervisory model to generate a new corresponding high-dimensional feature representation; Step 3: Build a multi-task meta-learning optimization model for the power market, and use high-dimensional feature representation to train the multi-task meta-learning optimization model for the power market, so that the multi-task meta-learning optimization model for the power market can adapt to the dynamic changes of the power market and optimize various market indicators; specifically: S301, construct multiple tasks according to different scenarios in the power market, and use different task-specific sub-feature sets from the high-dimensional feature representation as task data; S302, combining task-specific loss terms and innovative inter-task coordination regularization terms to construct a multi-objective loss function of a multi-task meta-learning optimization model, ensuring that the multi-task meta-learning optimization model can adapt to and optimize a variety of market indicators in different task environments; wherein the multi-objective loss function of the multi-task meta-learning optimization model It is expressed as follows: Where N represents the total number of tasks; Represents task T i The specific loss, W i It is task T i Parameters; α i represents the weight of the task loss, which is used to balance the impact of different tasks; λ represents the task-specific regularization weight; Represents task i for feature H ij The loss gradient is used to evaluate the importance of features; ρ j represents the importance weight of feature j; β represents the weight of the inter-task coordination regularization term; γ jk represents the similarity weight between task j and task k, which is used to control the coordination of task parameters; S303, jointly training the multi-task meta-learning optimization model through multiple tasks, so that meta-parameters are shared on different tasks, thereby adapting to new tasks; S304, verifying and optimizing the trained multi-task meta-learning optimization model; Step 4: Build and dynamically adjust the real-time operation indicator system of the power market; specifically include: S401, constructing a feature mapping function according to meta parameters and real-time data streams to generate real-time market indicators; S402, introducing an environment change response mechanism into the real-time data stream to detect changes in the real-time data stream and adaptively adjust indicators; S403, designing an adaptive feedback optimization mechanism to update the index mapping matrix and the parameters of the feature mapping function in the index multi-task meta-learning optimization model, and finally constructing a total loss function; Step 5: Construct a multi-objective strategy optimization algorithm to design the optimal market strategy based on the real-time adjusted market indicators; specifically, it includes: According to the market status and the initial strategy set at the time of historical data generation, a multi-objective optimization function is constructed based on the initial strategy set combined with the weighted sum of multiple market objectives, and the strategy set is iteratively updated using the optimization algorithm according to the multi-objective optimization function to obtain the optimal market strategy, and a strategy evaluation function is constructed to evaluate the optimal market strategy, and the weights in the optimization objective function are dynamically adjusted according to the evaluation results to dynamically update the market strategy; Step 6: Perform system learning and adaptive updates based on the optimal market strategy to maintain optimal performance.

2. The method for constructing an indicator system for power market operation according to claim 1, characterized in that: The preprocessing includes aligning the time series using a dynamic time correction method; The Step 1 further includes: For time-sequence data, a sliding window is used to calculate local features; for unstructured data, an embedded semantic transformation method is used to extract useful information; A dynamic missing value filling method based on multivariate interpolation is used to fill missing values. At the same time, an anomaly detection method based on adaptive threshold is used to dynamically set the detection threshold according to the local statistical characteristics of the data. If the data is greater than the detection threshold, the current data is marked as an outlier.

3. The method for constructing an indicator system for power market operation according to claim 1, characterized in that: In the self-supervised loss function The time difference constraint is also introduced in the self-supervisory loss function, and the self-supervised loss function is expressed as follows: in, Represents the target matrix predicted by the model, G i represents the actual target matrix, λ represents the weight of the time difference constraint, It is used to measure the prediction changes between consecutive time steps, ensuring that the model not only learns the point-to-point prediction ability of the data, but also captures the smoothness and consistency in time; n represents the number of samples; t represents the time step; t' represents the total time step; The self-supervisory model based on a multi-layer bidirectional long short-term memory network is constructed according to the self-supervisory loss function, specifically including: Randomly initialize the weight parameters of the self-supervised model based on a multi-layer bidirectional long short-term memory network; The forward propagation function of the multi-layer bidirectional long short-term memory network is used to calculate the predicted output of each sample; Calculate the gradient according to the self-supervised loss function, and then use the back-propagation algorithm to update the weight parameters of the self-supervised model; Through multiple rounds of iterations, minimize the self-supervised loss function Until convergence.

4. The method for constructing an indicator system for power market operation according to claim 1, characterized in that: In S301, a task diversity factor is added when constructing multiple tasks to ensure that the input features of multiple task subsets remain different; The S303 specifically includes: Randomly initialize the meta-parameters of the multi-task meta-learning optimization model and set the initial learning rate and meta-learning rate; For each meta-learning iteration, randomly select a task from the task set, randomly sample a mini-batch of data from the task-specific subset of inputs, calculate the task loss, and update the task-specific parameters via gradient descent; After all tasks are updated, the meta-loss of all tasks is calculated and the meta-parameters are updated by gradient descent. Repeat the above steps until the meta-parameters converge or meet the preset optimization conditions.

5. The method for constructing an indicator system for power market operation according to claim 4, characterized in that: The S304 specifically includes: Using a set of validation tasks Apply the trained meta-parameters W meta , calculate the validation loss for each task Weight the task loss according to the validation loss α i , feature importance weight ρ j and task similarity weight γ jk Perform dynamic adjustments and train again to optimize model performance, as shown below: in, represents the adjusted task loss weight; represents the adjusted feature importance weight; represents the adjusted task similarity weight; corr(H j ,H k ) represents the feature set H j and H k The correlation coefficient between .

6. A method for constructing an indicator system for power market operation according to claim 1, characterized in that: The S401 specifically includes: Design a feature mapping function to convert the real-time data stream into a high-dimensional feature space, and use the indicator mapping matrix extracted from the multi-task meta-learning optimization model to map the feature vector mapped to the high-dimensional space to the indicator space to generate the initial market indicator vector; Among them, the environmental change response mechanism senses the intensity of changes in the market environment by calculating the change rate of real-time data. If the change rate of real-time data exceeds a preset threshold, it is considered that the market environment has changed significantly and the indicator system needs to be adjusted.

7. A method for constructing an indicator system for power market operation according to claim 6, characterized in that: The total loss function Including indicator adjustment loss function and the regularization term Ω)t_, expressed as follows: in, represents the stability control term, which is used to prevent instability caused by frequent adjustments, and γ represents the weight factor of stability control; The indicator adjusts the loss function Contains real-time indicator errors, expressed as follows: Among them, M i )t_ represents the i-th market indicator currently calculated; represents the target i-th market indicator, which is derived based on historical data and forecasting models; λ1 represents the time smoothing regularization weight; m represents the number of features; Δt represents the time change.

8. The method for constructing an indicator system for power market operation according to claim 1, characterized in that: The multi-objective optimization function It is expressed as follows: Where n represents the dimension of the strategy set; g k (U(t),S(t)) represents the kth objective function associated with strategy U(t) and market state S(t); w k is the weight of the kth objective function, which is used to balance the importance of different objectives.

9. The method for constructing an indicator system for power market operation according to claim 1, characterized in that: The system learning and system adaptive updating according to the optimal market strategy specifically include: S601, construct an adaptive memory module to store past strategy data and market status information, the adaptive memory module includes a dynamically updated memory matrix M mem (t)∈R m×k , used to save the market state vector S(t) and the corresponding strategy U(t) of the most recent k times. The update of the memory matrix is ​​based on a rolling window mechanism. Each time it is updated, the earliest set of data is removed and the latest data is added. S602: Use the current market strategy U(t+1) and the historical information in the adaptive memory module to construct a new adaptive loss function At the same time, the difference between the current market status and historical data is combined to minimize the error between the prediction of the multi-task meta-learning optimization model and the actual market; the adaptive loss function It is expressed as follows: in, is the j-th market state value predicted by the model; S j (t+1) is the actual j-th market state value; γ is a weight factor that controls the impact of past market state differences; i is the index in the past k times, which is used to calculate the weight of past data; S603: Based on the strategy evaluation results and the historical strategy data of the adaptive memory module, a strategy update function φ(·) is constructed to adjust the strategy in real time according to market changes, as shown below: Among them, U new (t+1) represents the updated strategy vector; β represents the strategy adjustment step size, which is used to control the intensity of strategy update; Represents the gradient of the policy evaluation function, which is used to adjust the policy; S604. Repeatedly adjust and optimize strategies and system parameters to maintain the adaptability and robustness of the system.

10. A system for constructing an index system for power market operation, characterized in that: The system comprises the following units: The power market data collection unit is used to integrate data from different sources in the power market and pre-process the data; A data self-supervision unit is used to obtain pre-processed data and perform self-supervision data extraction from the pre-processed data; Specifically include: S201, constructing a feature matrix and a target matrix according to the preprocessed data set, wherein the feature matrix is ​​extracted from the preprocessed data set, and the target matrix is ​​generated by a sliding window technique, and is used to predict the market status in the future t time steps; S202, constructing a self-supervised loss function according to the error of future state prediction, and then constructing a self-supervised model based on a multi-layer bidirectional long short-term memory network according to the self-supervised loss function, and using a gradient descent algorithm to optimize the model parameters of the self-supervised model; S203, using the trained self-supervisory model to generate a new corresponding high-dimensional feature representation; The data meta-learning unit is used to build a multi-task meta-learning optimization model for the power market and use high-dimensional feature representation to train the multi-task meta-learning optimization model for the power market, so that the multi-task meta-learning optimization model for the power market can adapt to the dynamic changes of the power market and optimize multiple market indicators; specifically, it includes: S301, construct multiple tasks according to different scenarios in the power market, and use different task-specific sub-feature sets from the high-dimensional feature representation as task data; S302, combining task-specific loss terms and innovative inter-task coordination regularization terms to construct a multi-objective loss function of a multi-task meta-learning optimization model, ensuring that the multi-task meta-learning optimization model can adapt to and optimize a variety of market indicators in different task environments; wherein the multi-objective loss function of the multi-task meta-learning optimization model It is expressed as follows: Where N represents the total number of tasks; Represents task T i The specific loss, W i It is task T i Parameters; α i represents the weight of the task loss, which is used to balance the impact of different tasks; λ represents the task-specific regularization weight; Represents task i for feature H ij The loss gradient is used to evaluate the importance of features; ρ j represents the importance weight of feature j; β represents the weight of the inter-task coordination regularization term; γ jk represents the similarity weight between task j and task k, which is used to control the coordination of task parameters; S303, jointly training the multi-task meta-learning optimization model through multiple tasks, so that meta-parameters are shared on different tasks, thereby adapting to new tasks; S304, verifying and optimizing the trained multi-task meta-learning optimization model; The data indicator system construction unit is used to construct and dynamically adjust the real-time operation indicator system of the power market; specifically includes: S401, constructing a feature mapping function according to meta parameters and real-time data streams to generate real-time market indicators; S402, introducing an environment change response mechanism into the real-time data stream to detect changes in the real-time data stream and adaptively adjust indicators; S403, designing an adaptive feedback optimization mechanism to update the index mapping matrix and the parameters of the feature mapping function in the index multi-task meta-learning optimization model, and finally constructing a total loss function; The indicator system optimization unit is used to construct a multi-objective strategy optimization algorithm and design the optimal market strategy based on the market indicators adjusted in real time; specifically, it includes: According to the market status and the initial strategy set at the time of historical data generation, a multi-objective optimization function is constructed based on the initial strategy set combined with the weighted sum of multiple market objectives, and the strategy set is iteratively updated using the optimization algorithm according to the multi-objective optimization function to obtain the optimal market strategy, and a strategy evaluation function is constructed to evaluate the optimal market strategy, and the weights in the optimization objective function are dynamically adjusted according to the evaluation results to dynamically update the market strategy; The system optimization unit is used to perform system learning and system adaptive updates based on the optimal market strategy to maintain optimal performance.

Images