Trusted evaluation method and model for power grid dispatching algorithm, terminal equipment and storage medium

Abstract

The invention discloses a credibility evaluation method and model for a power grid dispatching algorithm, terminal equipment and a storage medium, and the method comprises the steps: collecting a power grid data set, carrying out the data cleaning and normalization processing of the data set, and dividing the data set into a training set, a test set and a verification set; respectively training and verifying the power grid dispatching algorithm by adopting the training set and the verification set to obtain a trained power grid dispatching algorithm; inputting a test set into the trained power grid dispatching algorithm for prediction to obtain a preliminary predicted value, and carrying out reverse normalization processing on the preliminary predicted value to obtain a predicted value; calculating an error based on the predicted value; a membership function is designed and is used for carrying out fuzzy processing on the error to obtain a membership function value; and carrying out defuzzification processing on the membership function value to obtain the credibility of the power grid dispatching algorithm, thereby ensuring that the AI algorithm has high credibility and reliability in the application of the power grid dispatching.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence, power system management and optimized dispatching, and in particular relates to a trusted evaluation method, model, terminal device and storage medium for a power grid dispatching algorithm. Background Art

[0002] Artificial intelligence (AI) technology offers new solutions for power grid dispatching. However, the complexity and uncertainty of AI algorithms make their credibility in practical applications a pressing issue. Therefore, research on trustworthiness metrics and evaluation methods for AI algorithms in power grid dispatching has both important theoretical value and significant practical significance.

[0003] First, the application of AI technology in power grid dispatching can significantly improve the efficiency and intelligence of dispatching. Through machine learning and deep learning algorithms, dispatch centers can uncover underlying patterns in massive amounts of historical data and optimize the allocation and use of power resources. However, the application of AI algorithms in power grid dispatching is still in its early stages, and their performance in different scenarios is highly uncertain. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the uncertainty of AI algorithms in different power grid dispatching scenarios and construct a trustworthy evaluation model for evaluating the credibility of power grid dispatching AI algorithms in different scenarios. The present invention provides a trustworthy evaluation method for power grid dispatching AI algorithms in prediction and optimization dispatching scenarios, comprising the following steps:

[0005] Collecting a power grid data set, performing data cleaning and normalization on the data set, and dividing the data set into a training set, a test set, and a validation set;

[0006] The training set and validation set are used to train and validate the power grid dispatch AI algorithm respectively, and a trained power grid dispatch AI algorithm is obtained;

[0007] Inputting the test set into the trained power grid dispatch AI algorithm for prediction to obtain a preliminary prediction value, and denormalizing the preliminary prediction value to obtain a prediction value;

[0008] calculating an error based on the predicted value;

[0009] Designing a membership function for fuzzifying the error to obtain a membership function value;

[0010] The membership function value is defuzzified to obtain the credibility of the power grid dispatching AI algorithm and verify the effectiveness of the power grid dispatching AI algorithm. The credibility evaluation method of the present invention is strictly carried out in accordance with scientific evaluation standards and methods to ensure the reliability and practicality of the evaluation model.

[0011] Optionally, the data cleaning includes at least one of missing value processing, outlier processing and inconsistent data processing.

[0012] Optionally, the data set includes n sets of input data P and target data T; the input data refers to data input into the power grid dispatching algorithm to be trained, and the target data refers to data compared with predicted data in the power grid dispatching AI algorithm to be trained during the training process;

[0013] The normalization process includes: The input data is normalized by the following calculation formula: ; is the maximum value of the input data after data cleaning, is the minimum value of the input data after data cleaning, After normalization, Group input data, After data cleaning Group input data, The value range is 1 to ; The target data is normalized by the following calculation formula: ;

[0014] Target data after data cleaning The maximum value in Target data after data cleaning The minimum value in is the i-th group of target data after data cleaning, is the i-th group of target data after normalization;

[0015] Optionally, the preliminary predicted value is denormalized to obtain a predicted value, specifically:

[0016]

[0017] is the maximum value among the preliminary predicted values, is the minimum value among the preliminary predicted values; represents the predicted value, Represents the i-th preliminary prediction value.

[0018] Optionally, the power grid scheduling AI algorithm is a transformer model or an optimization scheduling algorithm based on stacking integration.

[0019] Optionally, the stacking-based optimization scheduling algorithm specifically integrates the deep Q network DQN, the deep deterministic policy gradient algorithm DDPG, and the asynchronous advantage action evaluation algorithm A3C into one algorithm using stacking.

[0020] Optionally, the membership function designed in step 5 is:

[0021]

[0022]

[0023]

[0024] in, Indicates error, 、 、 Both represent the set thresholds; represents the low fuzzy set membership function, Represents the fuzzy set membership function, Represents the highly fuzzy set membership function.

[0025] Optionally, the error includes a mean square error , mean absolute error , root mean square error ;

[0026] ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 };

[0027] 、 、 is to calculate the mean square error When the membership function value is , the threshold value is set;

[0028] 、 、 is to calculate the mean absolute error When the membership function value is , the threshold value is set;

[0029] 、 、 is to calculate the root mean square error When the membership function value is , the threshold is set.

[0030] Optionally, in step 6, the membership function value is defuzzified to obtain the credibility of the power grid dispatching algorithm. , specifically:

[0031]

[0032] in, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

[0033] Optionally, based on the credibility Categorize credibility into high, medium and low:

[0034] When credibility When the value of is greater than 0.7 and less than or equal to 1, the credibility is high;

[0035] When credibility When the value of is greater than 0.4 and less than or equal to 0.7, the credibility is medium;

[0036] Credibility When the value of is greater than 0 and less than or equal to 0.4, the credibility is low.

[0037] Optionally, training and testing the stacking-based optimization scheduling algorithm includes the following steps:

[0038] Step 1, the training set is ,in is the feature vector of the training set of power grid data, is the label value of the training set;

[0039] Test set .

[0040] Step 2: The training set Divide into 5 non-overlapping training subsets .

[0041] Step 3: For each base learner model among the DQN, DDPG and A3C base learner models, select One of them is the test set, and the other 4 are training sets for learning. Repeat this step until the 5 training subsets are cycled through, realizing the Stacking transformation of the training set by the base learning model. Convert to prediction result output , that is, the training set After the first stage of stacking, it is converted into a meta-training set , the meta-training set for .

[0042] Step 4: Prediction results The five columns in the , as the meta-training set ;

[0043] Step 5: The meta-training set after the first stage of stacking and the meta-training set Input the meta-learning model A3C for training, and then input the meta-test set into the meta-learning model A3C Make predictions and get the trained model A3C.

[0044] The present invention also provides a credible evaluation device for a power grid dispatching algorithm, comprising:

[0045] A data processing module is used to collect power grid data sets, perform data cleaning and normalization on the data sets, and divide the data sets into training sets, test sets, and validation sets;

[0046] A training module is used to train and verify the power grid dispatching algorithm using a training set and a verification set, respectively, to obtain a trained power grid dispatching algorithm;

[0047] A prediction module is used to input the test set into the trained power grid dispatching algorithm to perform prediction, obtain a preliminary prediction value, and perform a denormalization process on the preliminary prediction value to obtain a prediction value;

[0048] A calculation module, configured to calculate an error based on the predicted value;

[0049] A fuzzification module is used to design a membership function, and is used to fuzzify the error to obtain a membership function value;

[0050] The defuzzification module is used to defuzzify the membership function value to obtain the credibility of the power grid scheduling algorithm.

[0051] Optionally, in the fuzzification module, the membership function is designed as:

[0052]

[0053]

[0054]

[0055] in, Indicates error, 、 、 Both represent the set thresholds; represents the low fuzzy set membership function, Represents the fuzzy set membership function, Represents the highly fuzzy set membership function.

[0056] Optionally, in the membership function in the fuzzification module, the error includes the mean square error , mean absolute error , root mean square error ;

[0057] ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 };

[0058] 、 、 is to calculate the mean square error When the membership function value is , the threshold value is set;

[0059] 、 、 is to calculate the mean absolute error When the membership function value is , the threshold value is set;

[0060] 、 、 is to calculate the root mean square error When the membership function value is , the threshold is set.

[0061] Optionally, in a defuzzification module, the membership function value is defuzzified to obtain the credibility of the power grid dispatching algorithm; wherein the defuzzification process is specifically:

[0062]

[0063] in, represents the credibility of the power grid dispatch algorithm, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

[0064] The present invention also provides a terminal device comprising a processor and a memory, the memory storing a plurality of instructions; the processor loads instructions from the memory to execute the steps of the trustworthiness assessment method for a power grid dispatching algorithm of the present invention. The present invention also provides a computer-readable storage medium storing a plurality of instructions suitable for loading by the processor to execute the steps of the trustworthiness assessment method for a power grid dispatching algorithm of the present invention. Advantageous Effects: First, input data is preprocessed, including data cleaning and normalization, to ensure its integrity and consistency. By eliminating missing values and outliers and unifying data dimensions, subsequent model training and evaluation are more accurate and reliable. High-quality data is fundamental to ensuring the accuracy of analytical models. The present invention's strict requirements in the data preprocessing stage effectively improve the reliability of the entire evaluation process. Next, reliable prediction models and optimization scheduling models are constructed for prediction and optimization scheduling scenarios, respectively. Through multi-factor fuzzification processing and a fuzzy inference system, prediction errors and other relevant factors are converted into fuzzy credibility. This method comprehensively considers the influence of multiple factors, making the evaluation results more comprehensive and accurate, avoiding the potential bias caused by single-metric evaluation, and improving the scientific nature and credibility of the evaluation model. Finally, a defuzzification method is used to calculate a specific credibility value, making the evaluation results specific and quantifiable. This step ensures that the model evaluation results not only have the inclusiveness of fuzzy theory but also have specific numerical representations, making the evaluation results more intuitive and easy to understand. Accurate credibility calculation provides strong support for the performance evaluation of power grid dispatch AI algorithms.

[0065] In summary, the present invention provides a comprehensive and reliable method for trustworthy evaluation of power grid dispatching AI algorithms for power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of the trustworthy evaluation method of power grid dispatch AI algorithm in prediction scenarios;

[0067] Figure 2 This is a credibility analysis chart of Transformer on different data amounts;

[0068] Figure 3 This is a credibility analysis chart of Transformer on different data qualities;

[0069] Figure 4 This is a flow chart of the trustworthy evaluation method of power grid dispatch AI algorithm in the optimization dispatch scenario;

[0070] Figure 5 It is a credibility graph comparing different models on the same training round;

[0071] Figure 6 It is a credibility graph comparing different models on the same sample size. DETAILED DESCRIPTION

[0072] The present application is further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0073] Example 1

[0074] To ensure the safe and stable operation of the power grid, it is necessary to conduct rigorous credibility assessments on grid dispatch AI algorithms. By constructing a set of scientific and reasonable credibility metrics, we can comprehensively evaluate the performance of AI algorithms in different dispatch scenarios, thus providing a basis for algorithm optimization and improvement.

[0075] Secondly, studying the credibility of AI algorithms for power grid dispatch can effectively reduce the risks associated with algorithmic errors. By establishing a dynamic assessment method, the operational status of AI algorithms for power grid dispatch can be monitored in real time, potential issues can be identified and corrected promptly, and the safety and reliability of power grid dispatch can be ensured.

[0076] In addition, by conducting credibility assessment on the grid dispatching AI algorithm, the accuracy of its prediction and dispatching can be ensured, thereby maximizing energy utilization efficiency. This paper proposes a credibility assessment method for grid dispatching AI algorithm in prediction scenarios. Figure 1 As shown, the following steps are included:

[0077] Step 1: Collect the power grid data set, clean and normalize the data set, and divide the data set into training set, test set and validation set;

[0078] Collect power grid data sets, specifically including the actual load conditions of the power grid at different time periods, usually time series data, weather data, power generation data, etc. The data in the data sets are cleaned, normalized, and partitioned to form a high-quality multi-dimensional data set;

[0079] Step 1.1: Clean the data set to ensure there are no missing values or outliers.

[0080] (1) Missing value processing

[0081] In the actual process of acquiring information and data, various reasons can lead to data loss and gaps. Different methods are used to address these missing values based on the variable's distribution characteristics and importance. If a variable has a high missing rate (greater than 80%), low coverage, and low importance, the variable can be directly deleted. This method is called variable deletion. If the missing rate is low (less than 95%) and the importance is low, basic statistics (maximum, minimum, mean, median, mode) are used to fill in the gaps based on the data distribution. This method is called missing value filling. For missing data, the decision to "delete" or "fill" is generally based on the missing rate.

[0082] When collecting data, there may be reasons such as machine failure that cause the data at that time to not be recorded normally, resulting in incomplete data. The integrated energy system data is periodic, that is, in different years, but on the same date, the data between the two are similar. Therefore, when there are missing values, you can find data from two years before and after the date to supplement it. There are three ways to calculate it: 1) When the day with a missing value is a working day, and the corresponding day in the previous year is a non-working day, find the working day data closest to the date to supplement it; 2) When the day with a missing value is a holiday, and the corresponding day in the previous year is a working day, find the non-working day data closest to the date to supplement it; 3) When the type of the day with a missing value is the same as the date type corresponding to the previous year, it can be directly replaced with the data from the previous year. The specific formula is as follows:

[0083]

[0084] Where: is the number of similar days; is the data value of the similar day, Indicates the date, Indicates time.

[0085] (2) Outlier processing

[0086] Outliers are the norm in data distribution. Data outside a specific distribution area or range is usually defined as anomaly or noise. A common approach is to remove outliers.

[0087] Interference from factors such as power equipment failures can cause the final measurement results to deviate from the normal value. These points are called outliers. Therefore, it is necessary to identify and process outliers. There are two common methods for correcting outliers: horizontal processing and vertical processing.

[0088] a) Horizontal processing method

[0089] Typically, data from integrated energy systems such as wind power, photovoltaic power, and electric loads appear as a smooth curve over a continuous period of time. Under normal circumstances, these data do not change dramatically over a short period of time. In other words, the difference between the data at one moment and the next is not too large. If the data change threshold is exceeded, the data collected at that moment is considered an outlier and requires correction. The formula is as follows.

[0090] If the data collected at time t appears:

[0091] m a x [ | P t + 1 − P t | , | P t − P t − 1 | ] > e t

[0092] but If it is an abnormal value, it needs to be corrected, where is the data collected at time t+1, is the data collected at time t, is the data collected at time t-1, It is to set the threshold of data change.

[0093] First, collect 5 data immediately adjacent to time t and calculate their average value:

[0094]

[0095] Then, collect three data immediately adjacent to time t and calculate their average value:

[0096]

[0097] Next, collect two data points immediately adjacent to time t and calculate their average value:

[0098]

[0099] Finally, the three average values are weighted to obtain the final data correction value. :

[0100]

[0101] Where: is the weighted value.

[0102] b) Vertical processing method

[0103] Generally speaking, comprehensive energy system data such as wind power, photovoltaic power and power load data are all periodic. If the data of the same day type factors and at the same time are similar, then the difference between them is within the set change threshold. If the difference is too large and exceeds the change threshold, the data at this moment is theoretically an outlier and needs to be corrected.

[0104] If the data collected at time t appears:

[0105]

[0106] but If it is an abnormal value, it needs to be corrected. Get the data correction value .

[0107]

[0108] Where: is the average value of the same day type factor at time t over the past few days, is the load data change threshold set, is the data collected at time t.

[0109] (3) Inconsistent data processing

[0110] During the actual data production process, due to human factors or other reasons, the recorded data may contain inconsistencies. These inconsistent data need to be cleaned up before analysis. For example, errors during data entry can be corrected by comparing them with the original records, and knowledge engineering tools can also be used to detect data that violates rules.

[0111] In step 1.2, normalize the data in the dataset to a uniform dimension, and then divide it into training, test, and validation sets. Data processing accounts for over 70% of the time spent on the entire data analysis project. Therefore, the quality of the data directly determines the accuracy of the analysis model.

[0112] Demand-side flexible resources are connected to the grid in the form of aggregates, providing adjustable resources. The inputs to the aggregate response model include date information, such as month and day; electricity price information, such as marginal price, generation cost, congestion cost, and network loss cost; and weather information, such as temperature and humidity. Different input data have different dimensions, so normalization is required to ensure accurate and efficient model training.

[0113] Prediction models require a large amount of input data, and the units and values of different data vary significantly. To improve calculation speed, the input data can be normalized. Normalizing the input data to a value between [0, 1] reduces the amount of input data required for calculation and increases calculation speed. The output result obtained from normalized data is a normalized value. As a preliminary prediction value, it cannot be used directly as the true value. It must be denormalized to obtain the true output value.

[0114] The data set includes n groups of input data P and target data T; input data refers to the data input into the power grid dispatching algorithm to be trained, and target data refers to the number compared with the predicted data in the power grid dispatching algorithm to be trained during the training process.

[0115] Normalize the data in the dataset. The specific process is as follows:

[0116] First, normalize the input data:

[0117]

[0118] Then normalize the target data:

[0119]

[0120] Finally, the predicted output is denormalized:

[0121]

[0122] Where, It is the maximum value of the input data after data cleaning; It is the minimum value of the input data after data cleaning; It is the maximum value of the target data after data cleaning; It is the minimum value of the target data after data cleaning; is the input data of group i after normalization; is the i-th group of input data after data cleaning, and the value of i ranges from 1 to n;;; is the i-th group of target data after data cleaning, is the i-th group of target data after normalization, is the maximum value among the preliminary predicted values, is the minimum value among the preliminary prediction values, represents the i-th preliminary prediction value, Represents the i-th predicted value, which is the same as the input data are of the same dimension.

[0123] Step 2: Use the training set and the validation set to train and validate the power grid dispatching algorithm respectively to obtain a trained power grid dispatching algorithm;

[0124] Step 3: Input the test set into the trained power grid dispatching algorithm for prediction to obtain a preliminary prediction value, and then perform a denormalization process on the preliminary prediction value to obtain a prediction value;

[0125] Step 4: Calculate the error based on the predicted value, including the mean square error , mean absolute error , root mean square error ;

[0126] Step 5: Design a membership function to fuzzify the error and obtain the membership function value;

[0127] The membership function is expressed as:

[0128]

[0129]

[0130]

[0131] in, Indicates error, 、 、 Both represent the set thresholds; represents the low fuzzy set membership function, Represents the fuzzy set membership function, Represents the highly fuzzy set membership function.

[0132] ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 };

[0133] 、 、 is to calculate the mean square error When the membership function value is , the threshold value is set;

[0134] 、 、 is to calculate the mean absolute error When the membership function value is , the threshold value is set;

[0135] 、 、 is to calculate the root mean square error When the membership function value is , the threshold is set.

[0136] Step 6: Defuzzify the membership function value to obtain the credibility of the power grid dispatch algorithm;

[0137]

[0138] in, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

[0139] Based on the credibility Categorize credibility into high, medium and low:

[0140] When credibility When the value of is greater than 0.7 and less than or equal to 1, the credibility is high;

[0141] When credibility When the value of is greater than 0.4 and less than or equal to 0.7, the credibility is medium;

[0142] Credibility When the value of is greater than 0 and less than or equal to 0.4, the credibility is low.

[0143] Example 2

[0144] This paper proposes a trustworthy evaluation method for power grid dispatching AI algorithm in prediction scenarios. Figure 1 As shown, the following steps are included:

[0145] Step 1: Collect a power grid data set, clean and normalize the data set, and divide the data set into a training set, a test set, and a validation set; the data cleaning, normalization, and data set division methods are the same as those in step 1 of Example 1.

[0146] Step 2: Use the training set and the validation set to train and validate the power grid dispatching algorithm respectively to obtain a trained power grid dispatching algorithm;

[0147] In this specific embodiment, the Transformer model is selected as the power grid scheduling algorithm to extract the time series features of input data and predict the data at the next time point. The Transformer model's core architecture is well-suited for processing time series data. The Transformer model can efficiently process long sequences of data and does not suffer from the vanishing or exploding gradient issues common with RNNs (recurrent neural networks) and LSTMs (long short-term memory networks). This makes it highly effective in handling power grid load forecasting tasks with long-term dependencies. The Transformer model utilizes a self-attention mechanism, allowing for parallel processing of the entire input sequence, rather than the incremental processing required by RNNs. This parallel computing capability significantly accelerates model training and prediction, making it suitable for processing large-scale power grid data and meeting real-time requirements. Furthermore, the Transformer model demonstrates superior prediction accuracy in many sequence prediction tasks, including natural language processing and time series forecasting. Its powerful feature extraction and representation capabilities enable it to provide highly accurate prediction results for power grid load forecasting, helping to improve the accuracy and reliability of power grid scheduling. Although deep learning models are often considered "black box" models, the Transformer model's self-attention mechanism provides a degree of interpretability. By analyzing attention weights, we can understand which input features the model focuses on when making predictions, which helps improve the transparency and credibility of the model.

[0148] The Transformer model consists of layer normalization, attention layer, and multilayer perceptron (MLP). The Transformer model can be represented as follows:

[0149]

[0150]

[0151] LN stands for Layer Normalization. Time series data is first normalized by LN. The time series data is then reshaped into feature vectors. These feature vectors, serving as the original input features, are then further normalized by batch normalization and linearly transformed to generate queries (Q), keys (K), and values (V). To further compress the temporal dimension and extract global information, the keys and values are average-pooled after generation. The query (Q) and the average-pooled keys (K) and values (V) are then fed into a multi-head self-attention (MHSA) mechanism to generate an attention output. The output of the self-attention mechanism is then added to the original input features via a residual connection. This approach not only preserves the original input information but also mitigates the vanishing gradient problem, making the model easier to train. These features are then reshaped again into a shape suitable for subsequent processing and fed into a fully-connected (MLP) layer for further extraction. Finally, the output of the fully-connected layer is added to its input features via a residual connection, completing the feature extraction process.

[0152] Input the training set and test set obtained in step 1 into the Transformer model, train, test, and verify the Transformer model to obtain a trained Transformer model.

[0153] Step 3: Input the test set into the trained Transformer model for prediction to obtain a preliminary prediction value, and perform denormalization on the preliminary prediction value to obtain a prediction value;

[0154] Step 4: Calculate the error based on the predicted value to evaluate the performance of the Transformer model and find its shortcomings.

[0155] Common error assessment metrics include:

[0156] Mean Squared Error (MSE):

[0157]

[0158] Mean Absolute Error (MAE):

[0159]

[0160] Root Mean Square Error (RMSE):

[0161]

[0162] Among them, for this specific embodiment, It represents the predicted value obtained after the Transformer model outputs the initial predicted value and then performs denormalization. is the number of samples input to the Transformer model. These error metrics serve as input factors for the subsequent multi-factor fuzzification process. Furthermore, the performance data of the prediction model provides the foundational data for the credibility assessment method of the present invention, enabling it to be evaluated based on actual prediction performance.

[0163] Step 5: Design a membership function to fuzzify the error and obtain the membership function value;

[0164] During the credibility assessment of predictive power grid dispatch AI algorithms, multiple influencing factors need to be fuzzified to provide input for the fuzzy inference system. The prediction errors considered primarily include mean squared error (MSE), mean absolute error (MAE), and root mean square error (RMSE). Fuzzification involves converting the mean squared error (MSE), mean absolute error (MAE), and root mean square error (RMSE) into fuzzy sets. This conversion of input factors into fuzzy sets is accomplished using fuzzy membership functions. The prediction error indicators (MSE, MAE, and RMSE) can be converted into fuzzy variables, which can be defined as "low," "medium," and "high" fuzzy sets.

[0165] For the mean square error MSE, the following is the specific membership function:

[0166]

[0167]

[0168]

[0169] in, represents the low fuzzy set membership function of MSE, represents the fuzzy set membership function of MSE, Represents the high fuzzy set membership function of MSE, 、 、 are pre-set thresholds, representing the boundaries of low, medium, and high fuzzy sets respectively. Represents the mean squared error value of the actual prediction of the model.

[0170] The same fuzzification method can be used for mean absolute error (MAE) and root mean square error (RMSE).

[0171] For the mean absolute error MAE fuzzification, the membership function used is expressed as:

[0172]

[0173]

[0174]

[0175] in, Represents the low fuzzy set membership function of MAE, represents the fuzzy set membership function of MAE, Represents the highly fuzzy set membership function of MAE, 、 、 are pre-set thresholds, representing the boundaries of low, medium, and high fuzzy sets respectively. Represents the mean absolute error of the actual predictions of the model.

[0176] For the RMSE fuzzification, the membership function used is expressed as:

[0177]

[0178]

[0179]

[0180] in, Represents the low fuzzy set membership function of RMSE, represents the fuzzy set membership function of RMSE, The high fuzzy set membership function representing RMSE, 、 、 are pre-set thresholds, representing the boundaries of low, medium, and high fuzzy sets respectively. Represents the mean absolute error of the actual predictions of the model.

[0181] Step 6: Defuzzify the membership function value to obtain the credibility of the power grid dispatching algorithm;

[0182] Through multi-factor fuzzy processing and fuzzy inference system, the mean square error (MSE), mean absolute error (MAE) and root mean square error (RMSE) are converted into fuzzy credibility based on mean square error, fuzzy credibility based on mean absolute error and fuzzy credibility based on root mean square error respectively.

[0183] The fuzzy inference system uses fuzzy logic rules to convert the fuzzified mean square error (MSE), mean absolute error (MAE), and root mean square error (RMSE) into fuzzy credibility outputs. The fuzzy inference system uses a set of fuzzy rules to infer the fuzzy credibility of the prediction model.

[0184] In existing technologies, fuzzy rules are usually determined by expert knowledge or experience, and they define the mapping relationship between input (fuzzified error index) and output (fuzzy credibility). Each fuzzy rule consists of a condition part and a conclusion part.

[0185] Rule 1: If the membership function value of the mean square error (MSE) is low (L), the membership function value of the MAE is low (L), and the membership function value of the RMSE is low (L), then the confidence level is high (H).

[0186] Rule 2: If the membership function value of MSE is medium (M), the membership function value of MAE is medium (M), and the membership function value of RMSE is medium (M), then the credibility is medium (M).

[0187] Rule 3: If the membership function value of MSE is high (H), the membership function value of MAE is high (H), and the membership function value of RMSE is high (H), then the credibility is low (L).

[0188] However, the above-mentioned method in the prior art may result in the same error with two membership function values, such as when When and Two values will result in two situations for the credibility of the final judgment;

[0189] Therefore, in this specific embodiment, a defuzzification method is used to calculate a specific credibility value and verify the effectiveness of the Transformer model.

[0190] The membership function value output by the fuzzy inference system is defuzzified to obtain a specific credibility value. The commonly used defuzzification method is the centroid method. The specific calculation formula is as follows:

[0191]

[0192] in, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

[0193] Through defuzzification processing, a specific credibility value is obtained Used to evaluate the credibility of AI prediction algorithms. The specific meanings are as follows:

[0194] High credibility (close to 1, for example, the credibility interval is: greater than 0.7 and less than or equal to 1): This indicates that the AI algorithm performs well in terms of accuracy, robustness, explainability, real-time and security, and is suitable for practical applications in power grid dispatching.

[0195] Medium credibility (credibility interval: greater than 0.4 and less than or equal to 0.7): This means that the AI algorithm performs well in some aspects, but may have shortcomings in other aspects and require further optimization and improvement.

[0196] Low credibility (credibility interval: greater than 0 and less than or equal to 0.4): This indicates that the AI algorithm performs poorly on multiple evaluation indicators and is not suitable for current power grid scheduling needs. It needs to be redesigned or another algorithm needs to be selected.

[0197] The specific credibility value obtained reflects the credibility of the prediction model in the current scenario. This value can be used as a key indicator to evaluate the accuracy and reliability of the prediction model, providing support for power grid scheduling decisions and operation management.

[0198]

[0199] Table 1 Comparison of credibility of different neural networks

[0200] To evaluate the reliability of AI algorithms for predictive power grid dispatch, a credibility assessment model was constructed. The Transformer model, LSTM, and MLP were subjected to credibility assessment. As shown in Table 1, the results show that the Transformer model exhibits the best performance in power grid dispatch, with a credibility of 0.85. This is due to its powerful feature extraction capabilities and advantages in processing time series data, particularly in multi-node short-term load forecasting. The LSTM model has a credibility of 0.5, indicating potential for application in power grid dispatch, but further optimization is required. The MLP model has a credibility of only 0.4, which presents significant limitations in power grid dispatch applications. For these classic predictive models, the credibility measurement model can effectively calculate the corresponding credibility values.

[0201] In order to verify whether the credibility evaluation method of the present invention can calculate the corresponding credibility of a specific prediction model if the disturbance of external factors causes the change of model performance, a set of experiments are designed, such as Figure 2As shown in the figure, four models were trained using the same dataset, with data sizes of 50,000, 100,000, 150,000, and 250,000, respectively. To assess the model's credibility, the Mean Sequential Error (MSE) metric was used, resulting in credibility scores of 0.4, 0.45, 0.6, and 0.8, respectively. Experimental results show that the credibility of the Transformer model increases significantly with increasing training data size.

[0202] like Figure 3 As shown in the figure, the Transformer model's performance under three different data quality conditions shows significant differences. On the original data, the Transformer model performs best, achieving a confidence score of 0.85. This indicates that the original data contains the most complete and accurate information, helping the model learn more accurate features. Adding noisy data causes model performance to degrade, with the confidence score dropping to 0.7. This is because noisy data interferes with the Transformer model's learning, causing it to learn incorrect patterns. The impact of missing data handling on the Transformer model's performance depends on the effectiveness of the missing data handling method. Effective handling methods, such as mean or median filling, can improve the confidence score to 0.8. However, inappropriate handling methods, such as directly removing missing data or using an incorrect filling method, can reduce the model's confidence. Therefore, despite the perturbations of data quality factors, the confidence metric model can still calculate the corresponding confidence value.

[0203] Example 3

[0204] This paper proposes a trustworthy evaluation method for power grid dispatching AI algorithm in prediction scenarios. Figure 4 As shown, the following steps are included:

[0205] Step 1: Collect a power grid data set, clean and normalize the data set, and divide the data set into a training set, a test set, and a validation set; the data cleaning, normalization, and data set division methods are the same as those in step 1 of Example 1.

[0206] The dataset includes grid topology and equipment information, historical load data, renewable energy generation data, market demand and price data, and relevant weather data. After cleaning and normalization, this data is used to build an optimization scheduling model and conduct error analysis to evaluate the credibility and performance of the AI algorithm in optimizing grid scheduling.

[0207] Step 2: Use the training set and the validation set to train and validate the power grid dispatching algorithm respectively to obtain a trained power grid dispatching algorithm;

[0208] Due to the limited performance of a single optimization scheduling algorithm, in order to improve the generalization and performance of the model, the power grid scheduling algorithm in this specific embodiment is an optimization scheduling algorithm based on stacking integration. The three optimization algorithms (Deep Q Network DQN, Deep Deterministic Policy Gradient Algorithm DDPG, and Asynchronous Dominance Action Evaluation Algorithm A3C) are integrated into one algorithm using stacking to fully utilize their respective advantages and make up for the shortcomings of a single algorithm. This stacking-based optimization scheduling algorithm adopts a two-layer model architecture: a base learner prediction model and a meta-prediction model. It aims to integrate the capabilities of multiple base learners to improve prediction performance.

[0209] Step 1: Divide the data set after data cleaning and normalization into the training set and the test set is ,in is the feature vector of the training set of power grid data, is the label value of the training set.

[0210] Step 2: In order to ensure the diversity of the training set, the training set Divide into 5 non-overlapping training subsets .

[0211] Step 3: Use 5 non-overlapping training subsets Perform 5-fold cross-validation on the base learner model DQN (BaseLearner DQN), base learner model DDPG (Base Learner DDPG) and base learner model A3C (BaseLearner A3C); select each of the DQN, DDPG and A3C base learner models Four of them are training sets, and the remaining one is a validation set. In this way, the three base learners will obtain five training models with different weights respectively, and then make predictions on the test set to obtain more robust prediction results. Until the five training subsets are cycled and predicted, the prediction results of the three base learners are extracted as output features, which are used as the output of the prediction model of the first layer of base learners to obtain the stacking transformation of the base learner model for the training set. The three base learners will respectively convert the five training subsets into Converted into 5 different training weights and predicted on the training set to get their corresponding prediction outputs , that is, the training set After the first stage of stacking, it is converted into a meta-training set , the meta-training set that is .

[0212] Step 4: In order to enhance the robustness of the meta-model, the prediction outputs of the three base learners are also Taking the average value ; The predicted outputs of the three base learner models are averaged to obtain the meta-training set after Stacking conversion , the meta-training set It means .

[0213] Step 5: The meta-training set after the first stage of stacking and the meta-training set They are respectively input as training sets into the meta-learning model A3C for five-fold cross-validation training and used as test sets. On the test set, the prediction outputs obtained are and Then the prediction output results obtained above are averaged to obtain the prediction results of the final meta-model .

[0214] In power grid dispatch optimization, the reward value can indicate the quality of the dispatch plan executed by the model, so the reward value is used as an evaluation indicator.

[0215] Among them: DQN inherits the idea of Q-Learning, uses Bootsraptex of Bellman formula, calculates the target value and uses a state-action value function Iterative optimization is performed with the goal of achieving convergence. Is a parameter A neural network that outputs all possible actions in the current state at each forward calculation. Then select the one with the largest The action with the value is taken as the optimal action, and the state is continuously updated in the next iteration. value until the value function converges.

[0216] J Q ( i ) = E s , a ∼ D [ 1 2 ( r ( s , a ) + c max a ′ ∈ A Q i − ( s ′ , a ′ ) − Q i ( s , a )) 2 ]

[0217] Where, is the reward value of the current action; is the action value function of the target network; is the main network action value function, is the discount factor, Represents expectations, i.e., the experience replay pool Medium sampling state and actions Expectations on.

[0218] In this specific embodiment, the state Indicates the current status or observation value of the environment, such as the current load of the system, power generation, equipment status, weather data and other information, actions Indicates that the status The decision or action below.

[0219] DQN uses a The greedy strategy is to The probability of the action value function is selected according to the input state and action value function ,by The probability of randomly selecting an action is . The entire training process is a linear change from large to small, so that the early stage of training focuses on exploring unknown space and the later stage of training focuses on using the optimal action. When collecting samples, DQN uses a first-in-first-out stack structure as an experience pool to store training sample data. , each time during training, a batch of samples is selected for gradient calculation and neural network parameter update.

[0220] In order to solve the control task of continuous action space, a parameter is added to the DQN algorithm. Policy Network , is a deterministic strategy, according to the input state Output a unique deterministic action , , represents a continuous action space. The Q network's target value update utilizes the temporal difference method with the bootstrap property of the Bellman equation, similar to the DQN algorithm. Since the action space is continuous, the policy network directly outputs state-action values, replacing the DQN algorithm's Q-value-based action optimization. The policy network acts as the Q network optimizer, aiming to output the maximum value of the current Q network. The updated gradients are derived entirely from the Q network, as shown below.

[0221] J Q ( i ) = E s , a ∼ D [ 1 2 ( r ( s , a ) + c Q i − ( s ′ , π ϕ − ( a ′ | s ′ )) − Q i ( s , a )) 2 ]

[0222] J π ( ϕ ) =− E s , a ∼ D [ Q i ( s , π ϕ ( a | s ))]

[0223] represents the loss function of the Q network, represents the loss function of the policy network.

[0224] DDPG uses an additive noise exploration approach. This means that the output of the policy network is equivalent to adding Gaussian noise with a mean of zero at the same latitude. The variance of the noise determines the intensity of exploration. This is equivalent to forming a Gaussian distribution centered around the current output action. Each update of the policy network shifts the output action toward the direction with a higher Q value in the Gaussian distribution, until all other directions within the distribution are worse. At this point, the policy network's output stabilizes near the optimal action, effectively balancing exploration and exploitation.

[0225] Both DQN and DDPG are off-policy algorithms. Both can reuse data from the experience pool for training, improving sample utilization. However, their training stability is poor, leading to overestimation in target value calculations and hindering gradient updates. On-policy algorithms use the Monte Carlo method to improve algorithm stability by randomly sampling multiple complete cycles of the latest policy to obtain an unbiased estimate of the current value function. Therefore, the representative A3C algorithm among on-policy algorithms was selected to compensate for this lack of stability.

[0226] Unlike the DDPG policy network that uses a deterministic policy, A3C uses a random policy and outputs the probability distribution of actions. It cannot update the gradient through grid calculation and needs to estimate the gradient through random sampling. Its advantage function Indicates Select action in status Advantage estimate:

[0227]

[0228] Where, For Status Start to The discounted cumulative return for a period of time at the end of the state; The number of state iterations shall not exceed the maximum number of iterations of an Episode; is the value function, the value network is responsible for evaluating the current state The “value” of the state is the expected return that can be obtained in this state. This network provides feedback to the policy network to guide the optimization of the policy.

[0229] The following is the value network The objective function.

[0230] J v ( i ) = E s , R ∼ r π [ 1 2 ( R − V i ( s )) 2 ]

[0231] The following is the strategic network Objective function, the policy network is responsible for generating the action distribution in each state and selecting the optimal action. The optimization goal is to increase the output probability of the policy network's advantage estimate of the larger action.

[0232] J π ( ϕ ) =− E s , a ∼ r π [log π ϕ ( a | s ) A ( s , a ) + oh H ( π ϕ ( a | s ))]

[0233] Where, For the strategy Sample distribution of the downsampling period; In state Select Action probability; In state Select Action advantage estimate; Is a hyperparameter that adjusts the relative weight of policy entropy loss and policy loss; is the policy entropy.

[0234] The stochastic policy employed by A3C inherently has exploratory properties. Actions are randomly sampled according to the probability distribution of the output actions. As training progresses, the action selection becomes increasingly optimal, and the randomness of the policy output gradually decreases, achieving a balance between exploration and exploitation. To prevent the policy from prematurely falling into a local optimum and becoming deterministic, the A3C algorithm incorporates a policy entropy loss to ensure the randomness of the policy network. As an on-policy algorithm, the model must use the latest policy for sampling during each update, and collected samples cannot be added to the experience pool for reuse. To ensure training stability, the A3C algorithm often employs parallel sampling to ensure sufficient samples for each update.

[0235] Step 3: Input the test set into the trained Transformer model for prediction to obtain a preliminary prediction value, and perform denormalization on the preliminary prediction value to obtain a prediction value.

[0236] Step 4: Calculate the error based on the predicted value. In this embodiment, the error includes mean square error (MSE), mean absolute error (MAE), and root mean square error (RMSE).

[0237] Step 5: Design a membership function to fuzzify the error and obtain the membership function value;

[0238] Through multi-factor fuzzy processing and fuzzy inference system, the mean square error (MSE), mean absolute error (MAE) and root mean square error (RMSE) are converted into fuzzy credibility based on mean square error, fuzzy credibility based on mean absolute error and fuzzy credibility based on root mean square error respectively.

[0239] The multi-factor fuzzy processing process is the same as that in the second embodiment, and both are implemented by using fuzzy membership function, except that the threshold is set. 、 、 Different from the second embodiment, ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 }.

[0240] Step 6: Defuzzify the membership function value to obtain the credibility of the power grid dispatching algorithm;

[0241] Fuzzy inference systems use fuzzy logic rules to convert fuzzified input factors into fuzzy confidence outputs. Fuzzy inference systems utilize a set of fuzzy rules to infer the fuzzy confidence of a predictive model. Fuzzy rules, typically determined by expert knowledge or experience, define the mapping between input (fuzzified error metrics) and output (fuzzy confidence).

[0242] In this embodiment, the output of the fuzzy inference system is defuzzified to obtain a specific credibility value. The commonly used defuzzification method is the centroid method. The specific calculation formula is as follows:

[0243]

[0244] in, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

[0245] Through defuzzification processing, a specific credibility value is obtained Used to evaluate the credibility of AI prediction algorithms. The specific meanings are as follows:

[0246] High credibility (greater than 0.7 and less than or equal to 1): indicates that the AI algorithm performs well in terms of accuracy, robustness, explainability, real-timeness, and security, and is suitable for practical applications in power grid dispatching.

[0247] Medium credibility (greater than 0.4 and less than or equal to 0.7): This means that the AI algorithm performs well in some aspects, but may have shortcomings in other aspects and requires further optimization and improvement.

[0248] Low credibility (greater than 0 and less than or equal to 0.4): This indicates that the AI algorithm performs poorly on multiple evaluation indicators and is not suitable for current power grid scheduling needs. It needs to be redesigned or another algorithm needs to be selected.

[0249] The specific credibility value obtained reflects the credibility of the prediction model in the current scenario. This value can be used as a key indicator to evaluate the accuracy and reliability of the prediction model, providing support for power grid scheduling decisions and operation management.

[0250] like Figure 5 As shown in the figure, the credibility of DQN, DDPG, A3C, and the ensemble model exhibits different trends at different training rounds. The ensemble model achieves the best credibility across all training rounds, demonstrating that the ensemble model combines the strengths of different algorithms to achieve better performance. DQN's credibility improves more slowly, with relatively weak performance in the early stages. However, its credibility gradually improves with increasing training rounds. This indicates that the DQN algorithm converges more slowly and requires more training rounds to achieve high performance. DDPG's credibility improves faster than DQN, showing good performance in the early stages and gradually approaching the optimal value as the number of training rounds increases. This indicates that DDPG converges relatively quickly and has certain advantages in solving continuous control problems. A3C's credibility improves the fastest, achieving high credibility in the early stages and quickly reaching the optimal value. This suggests that A3C converges the fastest, but its stability may be inferior to DDPG.

[0251] like Figure 6 To evaluate the credibility of different reinforcement learning algorithms for power grid scheduling, an IEEE 14-node system was used as the test environment. DQN, DDPG, A3C, and an integrated model were trained with 100, 200, 500, 1000, and 2000 training samples, respectively. The figure shows the credibility scores of different algorithms with different numbers of training samples.

[0252] As can be seen from the figure, the confidence scores of all algorithms increase with the number of training samples, indicating that more training data helps the model learn more accurate scheduling strategies. The ensemble model exhibits the best feasibility across all training sample sizes, demonstrating that it combines the strengths of different algorithms to achieve better performance.

[0253] Compared to other algorithms, DQN performs relatively poorly with fewer training samples. This is because it relies on an experience replay mechanism, requiring a large amount of sample data to learn an effective policy. DDPG's performance also suffers with fewer training samples. A3C, however, is less dependent on the number of samples due to its parallel learning mechanism, but it still benefits from more sample data.

[0254] Example 4

[0255] This paper proposes a trustworthy assessment model for power grid dispatching AI algorithms based on prediction and optimization scheduling, which includes the following modules:

[0256] A data processing module is used to collect power grid data sets, perform data cleaning and normalization on the data sets, and divide the data sets into training sets, test sets, and validation sets;

[0257] A training module is used to train and verify the power grid dispatching algorithm using a training set and a verification set, respectively, to obtain a trained power grid dispatching algorithm;

[0258] A prediction module is used to input the test set into the trained power grid dispatching algorithm to perform prediction, obtain a preliminary prediction value, and perform a denormalization process on the preliminary prediction value to obtain a prediction value;

[0259] A calculation module, configured to calculate an error based on the predicted value;

[0260] A fuzzification module is used to design a membership function, and is used to fuzzify the error to obtain a membership function value;

[0261] The defuzzification module is used to defuzzify the membership function value to obtain the credibility of the power grid scheduling algorithm.

[0262] In the fuzzification module, the membership function is designed as:

[0263]

[0264]

[0265]

[0266] in, Indicates error, 、 、 Both represent the set thresholds; represents the low fuzzy set membership function, Represents the fuzzy set membership function, Represents the highly fuzzy set membership function.

[0267] Among them, the error Including mean square error , mean absolute error , root mean square error ;

[0268] ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 }, ∈{ 、 、 };

[0269] 、 、 is to calculate the mean square error When the membership function value is , the threshold value is set;

[0270] 、 、 is to calculate the mean absolute error When the membership function value is , the threshold value is set;

[0271] 、 、 is to calculate the root mean square error When the membership function value is , the threshold is set.

[0272] In the defuzzification module, the membership function value is defuzzified using the following formula:

[0273]

[0274] in, represents the credibility of the power grid dispatch algorithm, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

[0275] This specific embodiment also provides a terminal device, including a processor and a memory, wherein the memory stores multiple instructions; the processor loads instructions from the memory to execute the steps in the trust evaluation method of the power grid scheduling algorithm in Example 1.

[0276] This specific embodiment also provides a computer-readable storage medium storing a plurality of instructions suitable for loading by a processor to execute the steps of the trust evaluation method for a power grid dispatching algorithm described in any one of the first embodiments. It should be noted that, with respect to the trust evaluation method for a power grid dispatching algorithm described in this application, those skilled in the art will appreciate that all or part of the process of implementing the trust evaluation method for a power grid dispatching algorithm described in this application can be accomplished by controlling related hardware through a computer program. The computer program can be stored in a computer-readable storage medium, such as a memory of a terminal device, and executed by at least one processor within the terminal device. During execution, the process may include the process of the embodiment of the vibration adjustment method described in this application. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), or the like. The present invention systematically analyzes the multi-dimensional hierarchical trust attributes of a power grid dispatching AI algorithm. First, by conducting an in-depth study of the structure and operating mechanism of the AI algorithm, the key factors affecting the algorithm's trustworthiness are identified, and a decomposition method is proposed. Second, the method studies how to obtain effective trust evidence from aspects such as data, methods, and operating environment, providing a foundation for the construction of the evaluation method. On this basis, we constructed a trustworthy evaluation method for AI algorithms applicable to typical scenarios such as prediction and optimization scheduling commonly used in power grid dispatch. Finally, we developed a dynamic evaluation model based on the association of trustworthy attributes to enable real-time evaluation of AI algorithms during operation, thereby improving the overall security and trustworthiness of power grid dispatch AI algorithms.

[0277] This paper proposes a trustworthy evaluation model and method for AI algorithms used in power grid dispatch in predictive and optimized dispatch scenarios. Combining deep learning, optimization algorithms, and smart grid technologies, the aim is to develop a systematic, quantitative evaluation model for assessing the trustworthiness of AI algorithms used in power grid dispatch, thereby improving the operational reliability and efficiency of power systems. Through data processing, model building and verification, evaluation indicator design, and trustworthiness calculation, this paper provides a reliable performance evaluation tool for the application of AI algorithms in power grid dispatch, thereby improving the operational reliability and efficiency of power systems.

[0278] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments are only for illustrating the technical ideas of the present invention and cannot be used to limit the scope of protection of the present invention. Any changes made on the basis of the technical solution in accordance with the technical ideas proposed by the present invention fall within the scope of protection of the present invention.

Claims

1. A trustworthy evaluation method for a power grid dispatching algorithm, characterized in that: The steps include: Collecting a power grid data set, performing data cleaning and normalization on the data set, and dividing the data set into a training set, a test set, and a validation set; The training set and validation set are used to train and validate the power grid dispatching algorithm respectively, and a trained power grid dispatching algorithm is obtained; Inputting the test set into the trained power grid dispatching algorithm for prediction to obtain a preliminary prediction value, and performing a denormalization process on the preliminary prediction value to obtain a prediction value; calculating an error based on the predicted value; Designing a membership function for fuzzifying the error to obtain a membership function value; The membership function value is defuzzified to obtain the credibility of the power grid dispatching algorithm.

2. The trust evaluation method for a power grid dispatching algorithm according to claim 1, characterized in that: The data cleaning includes at least one of missing value processing, outlier processing and inconsistent data processing.

3. The trust evaluation method for a power grid dispatching algorithm according to claim 1, characterized in that: The dataset includes Group input data and target data ; The input data refers to the data input into the power grid dispatching algorithm to be trained, and the target data refers to the number compared with the predicted data in the power grid dispatching algorithm to be trained during the training process; The normalization process includes: The input data is normalized by the following calculation formula: ； is the maximum value of the input data after data cleaning, is the minimum value of the input data after data cleaning, After normalization, Group input data, After data cleaning Group input data, The value range is 1 to ; The target data is normalized by the following calculation formula: ； For the target data after data cleaning The maximum value in Target data after data cleaning The minimum value in After data cleaning Group target data, After normalization, Group target data.

4. The trust evaluation method for a power grid dispatching algorithm according to claim 1, characterized in that: The preliminary predicted value is denormalized to obtain the predicted value, specifically: ； is the maximum value among the preliminary predicted values, is the minimum value among the preliminary predicted values; Indicates the Predicted values, Indicates the The preliminary forecast value.

5. The trust evaluation method for a power grid dispatching algorithm according to claim 1, characterized in that: The power grid dispatching algorithm is a transformer model or an optimization dispatching algorithm based on stacking integration.

6. The trust evaluation method for a power grid dispatching algorithm according to claim 5, characterized in that: The stacking-based optimization scheduling algorithm specifically integrates the deep Q network DQN, the deep deterministic policy gradient algorithm DDPG, and the asynchronous advantage action evaluation algorithm A3C into one algorithm using stacking.

7. The trustworthy evaluation method for a power grid dispatching algorithm according to claim 1, characterized in that: The membership function designed in step 5 is: ； ； ； in, Indicates error, 、 、 Both represent the set thresholds; represents the low fuzzy set membership function, Represents the fuzzy set membership function, Represents the highly fuzzy set membership function.

8. The trustworthy evaluation method for a power grid dispatching algorithm according to claim 7, characterized in that: The error includes the mean square error , mean absolute error , root mean square error ; ∈{ 、 、 }， ∈{ 、 、 }， ∈{ 、 、 }， ∈{ 、 、 }； 、 、 is to calculate the mean square error When the membership function value is , the threshold value is set; 、 、 is to calculate the mean absolute error When the membership function value is , the threshold value is set; 、 、 is to calculate the root mean square error When the membership function value is , the threshold is set.

9. The trustworthy evaluation method for a power grid dispatching algorithm according to claim 8, characterized in that: In step 6, the membership function value is defuzzified to obtain the credibility of the power grid dispatching algorithm. , specifically: ； in, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

10. The trustworthy evaluation method for a power grid dispatching algorithm according to claim 9, characterized in that: Based on the credibility Categorize credibility into high, medium and low: When credibility When the value of is greater than 0.7 and less than or equal to 1, the credibility is high; When credibility When the value of is greater than 0.4 and less than or equal to 0.7, the credibility is medium; Credibility When the value of is greater than 0 and less than or equal to 0.4, the credibility is low.

11. The trustworthy evaluation method for a power grid dispatching algorithm according to claim 6, characterized in that: The training and testing of the stacking-based optimization scheduling algorithm includes the following steps: The training set is ,in is the feature vector of the training set of power grid data, is the label value of the training set; Test set ; The training set Divide into 5 non-overlapping training subsets ; For each of the DQN, DDPG, and A3C base learner models, select One of them is the test set, and the other 4 are training sets for learning. Repeat this step until the 5 training subsets are cycled through, realizing the Stacking transformation of the base learner model to the training set. Convert to prediction result output ,make As a meta-training set ; The predicted results The five columns in the , as the meta-training set ; The meta-training set after the first stage of Stacking and the meta-training set Input the meta-learning model A3C for training, and then input the meta-test set into the meta-learning model A3C Perform the test and obtain the trained model A3C.

12. A trustworthy evaluation device for a power grid dispatching algorithm, characterized in that: include: A data processing module is used to collect power grid data sets, perform data cleaning and normalization on the data sets, and divide the data sets into training sets, test sets, and validation sets; A training module is used to train and verify the power grid dispatching algorithm using a training set and a verification set, respectively, to obtain a trained power grid dispatching algorithm; A prediction module is used to input the test set into the trained power grid dispatching algorithm to perform prediction, obtain a preliminary prediction value, and perform a denormalization process on the preliminary prediction value to obtain a prediction value; A calculation module, configured to calculate an error based on the predicted value; A fuzzification module is used to design a membership function, and is used to fuzzify the error to obtain a membership function value; The defuzzification module is used to defuzzify the membership function value to obtain the credibility of the power grid scheduling algorithm.

13. The trustworthy evaluation device for a power grid dispatching algorithm according to claim 12, characterized in that: The membership function is designed as: ； ； ； in, Indicates error, 、 、 Both represent the set thresholds; represents the low fuzzy set membership function, Represents the fuzzy set membership function, Represents the highly fuzzy set membership function.

14. The trustworthy evaluation device for a power grid dispatching algorithm according to claim 13, characterized in that: The error includes the mean square error , mean absolute error , root mean square error ; ∈{ 、 、 }， ∈{ 、 、 }， ∈{ 、 、 }， ∈{ 、 、 }； 、 、 is to calculate the mean square error When the membership function value is , the threshold value is set; 、 、 is to calculate the mean absolute error When the membership function value is , the threshold value is set; 、 、 is to calculate the root mean square error When the membership function value is , the threshold is set.

15. The trustworthy evaluation device for a power grid dispatching algorithm according to claim 14, characterized in that: In the defuzzification module, the membership function value is defuzzified to obtain the credibility of the power grid dispatching algorithm; the defuzzification process is specifically as follows: ； in, represents the credibility of the power grid dispatch algorithm, ∈{ 、 、 }, Mean square error The credibility obtained, Mean absolute error The credibility obtained, It represents the root mean square error Obtained credibility.

16. A terminal device, characterized in that: The system comprises a processor and a memory, wherein the memory stores a plurality of instructions; the processor loads instructions from the memory to execute the steps in the trustworthy evaluation method of the power grid dispatching algorithm according to any one of claims 1 to 12.

17. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps of the trustworthy assessment method of a power grid dispatching algorithm according to any one of claims 1 to 12.

Images