Power load change interpretable attribution analysis system, method and device and storage medium

Abstract

The invention discloses an interpretable attribution analysis system, method and equipment for power load change and a storage medium. The method comprises the following steps: performing weighted fusion on multi-site meteorological data and constructing a high-dimensional meteorological feature set containing nonlinear terms; decoupling the historical load sequence into a trend component and a periodic component; training a linear regression model by taking the high-dimensional meteorological feature set as input and the trend component as a target to obtain a regression coefficient of each feature; a control variable method is applied to calculate load trend variation caused by date types, meteorological factors and interaction changes between the analysis day and the comparison day; and generating a readable natural language attribution report. According to the method, the defects that an existing load prediction model is black-box-shaped and cannot quantitatively explain fluctuation causes are overcome, refined and interpretable attribution analysis of load change causes is achieved, and a direct and quantitative basis is provided for dispatching and decision making of an electric power system.

Description

Technical Field

[0001] This invention belongs to the field of power system load analysis technology, and particularly relates to an interpretable attribution analysis system, method, device and storage medium for power load changes. Background Technology

[0002] With the advancement of new power system construction, load characteristics are becoming increasingly complex. Accurate analysis of the causes of load changes is crucial for power grid dispatching, demand-side management, and renewable energy integration. Existing technologies mostly focus on load forecasting, while lacking effective means for ex-post attribution analysis of load changes. Traditional load analysis methods largely rely on the statistical patterns of historical load data, lacking refined modeling of the impact of external factors such as weather and holidays, making it difficult to explain the specific causes of load fluctuations.

[0003] In existing technologies, some methods use linear regression or machine learning models for load forecasting, but these methods have the following shortcomings:

[0004] Meteorological data processing is crude: it usually only uses meteorological data from a single station and does not consider regional meteorological differences or spatial weighting.

[0005] Characteristic engineering is simple: it uses only the original meteorological variables and does not introduce nonlinear terms (such as square terms or interaction terms) to capture complex relationships;

[0006] Inadequate handling of periodic loads: Failure to effectively separate the trend, periodicity, and residual components of the loads affects the accuracy of attribution.

[0007] Lack of interpretability: Most models are "black boxes" and cannot provide quantitative attribution formulas for load changes. Summary of the Invention

[0008] Objective of this invention: The objective of this invention is to provide an interpretable attribution analysis system and method for power load changes, enabling refined and quantitative analysis of the causes of power load changes, and ultimately generating interpretable attribution analysis formulas. Another objective of this invention is to provide an apparatus and storage medium for implementing the aforementioned interpretable attribution analysis method for power load changes.

[0009] Technical solution: The explanatory attribution analysis system for power load changes, characterized in that the system includes:

[0010] The data fusion and feature construction module constructs a regional high-dimensional meteorological feature set based on the fusion and nonlinear extension of multi-site meteorological data.

[0011] The load mechanism separation module is used to decouple historical load sequences into load trend components and load cycle components;

[0012] The attribution quantification analysis module includes a trained regression model used to establish a mapping between regional high-dimensional meteorological features and the load trend components, and to store the regression coefficients of each feature.

[0013] The controlled variable attribution engine is used to acquire high-dimensional meteorological characteristic data of the analysis day and the comparison day, call the attribution quantification analysis module, and use the controlled variable method to calculate the load change caused by the change of a single characteristic based on the regression coefficient.

[0014] An interpretable output generation module is used to summarize the load changes output by the control variable attribution engine, generate readable natural language attribution formulas, and output an analysis report.

[0015] Optionally, the data fusion and feature construction module, based on multi-site meteorological data fusion and nonlinear extension, includes:

[0016] The meteorological data are weighted according to the assigned meteorological weights to obtain comprehensive meteorological data;

[0017] The comprehensive meteorological data is used to generate square-term features and interaction-term features, which are then used as high-level meteorological features. The square-term feature refers to the product of any feature in the comprehensive meteorological data with itself, and the interaction-term feature refers to the product of any two different features in the comprehensive meteorological data. Through these square-term and interaction-term features, not only can the complex nonlinear physical relationship between load and meteorological features be fitted, but interpretability and accuracy can also be balanced.

[0018] Optionally, the load mechanism separation module includes:

[0019] The frequency domain transformation unit is used to transform the historical load sequence from the time domain to the frequency domain to obtain the spectral coefficients and frequency domain values ​​corresponding to the load data;

[0020] The filtering unit is used to construct a bandpass filter based on a preset target period, extract the load period component from the spectral coefficients, and convert it into a load period component in the time domain.

[0021] The trend calculation unit is used to calculate the load trend component based on the load data and the load cycle component.

[0022] Optionally, training the regression model includes:

[0023] A linear regression model is used, with the regional meteorological feature set as input and the load trend component as the training target, to fit the model. The complete load data is represented as the sum of the output of the linear regression model, the load cycle component, and a residual component (linear fitting error).

[0024] Optionally, the single feature change controlled by the control variable attribution engine includes one of the following: date type change, individual meteorological factor change, and high-order nonlinear feature change.

[0025] Optionally, the analysis report is in a structured format.

[0026] The explainable attribution analysis method for the changes in power load includes:

[0027] A regional high-dimensional meteorological feature set is constructed based on the fusion and nonlinear extension of multi-site meteorological data.

[0028] The historical load sequence is decoupled into load trend components and load cycle components;

[0029] Train a regression model to map the regional high-dimensional meteorological feature set to the load trend component and obtain the regression coefficients of each feature;

[0030] The characteristic data of the analysis day and the comparison day are input into the trained regression model, and the independent impact of the single characteristic change on the load trend is calculated based on the regression coefficients.

[0031] Summarize the independent impact of each feature, generate a readable natural language attribution formula, and output an analysis report.

[0032] Optionally, the multi-site meteorological data fusion and nonlinear extension includes:

[0033] Meteorological data from multiple stations in the region are weighted according to their assigned meteorological weights to obtain comprehensive meteorological data;

[0034] The comprehensive meteorological data is used to generate square-term features and interaction-term features, which are then used as advanced meteorological features. The square-term features refer to the product of any feature in the comprehensive meteorological data with itself, and the interaction-term features refer to the product of any two different features in the comprehensive meteorological data.

[0035] Optionally, mapping the regional high-dimensional meteorological feature set to the load trend component and obtaining the regression coefficients of each feature includes:

[0036] The historical load sequence is transformed from the time domain to the frequency domain to obtain the spectral coefficients and frequency domain values ​​corresponding to the load data;

[0037] A bandpass filter is constructed based on a preset target period, and the load period component is extracted from the spectral coefficients and converted into a time-domain load period component.

[0038] The load trend component is calculated based on the load data and the load cycle component.

[0039] The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements some or all of the steps in the above-described method for interpretable attribution analysis of power load changes.

[0040] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements some or all of the steps in the above-described method for interpretable attribution analysis of power load changes.

[0041] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0042] (1) It overcomes the deficiency of insufficient representativeness of data from a single meteorological station, and by introducing nonlinear features, the model can accurately characterize the complex marginal effects and coupling effects between meteorological factors (such as temperature and humidity) and load, thereby improving the accuracy and reliability of attribution analysis from the source.

[0043] (2) The trend component, which is mainly driven by meteorological factors, and the periodic component, which is mainly driven by date type, are effectively separated. This allows subsequent regression modeling to focus on the relationship between meteorological and trend components, eliminating the interference of periodic fluctuations on attribution analysis, thus making the attribution results for meteorological factors purer and more accurate.

[0044] (3) The inherent additivity of the linear regression model and the explicit physical meaning of its coefficients (i.e., the marginal contribution of features to the objective) make the model's decision-making process completely transparent. This overcomes the shortcomings of "black box" models such as neural networks that cannot be explained, and provides a solid mathematical foundation for the final quantitative attribution. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention;

[0046] Figure 2 This is a schematic diagram of the load sequence after decoupling in one embodiment of the present invention;

[0047] Figure 3 This is a comparison chart of the load curve fitting curves at a certain moment in one embodiment of the present invention;

[0048] Figure 4 This is a comparison chart of load curves for the analysis day and the control day in one embodiment of the present invention. Detailed Implementation

[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0050] An interpretable attribution analysis system for power load variations, characterized in that the system comprises:

[0051] The data fusion and feature construction module constructs a regional high-dimensional meteorological feature set based on the fusion and nonlinear extension of multi-site meteorological data.

[0052] The load mechanism separation module is used to decouple historical load sequences into load trend components and load cycle components;

[0053] The attribution quantification analysis module includes a trained regression model used to establish a mapping between regional high-dimensional meteorological features and the load trend components, and to store the regression coefficients of each feature.

[0054] The controlled variable attribution engine is used to acquire high-dimensional meteorological characteristic data of the analysis day and the comparison day, call the attribution quantification analysis module, and use the controlled variable method to calculate the load change caused by the change of a single characteristic based on the regression coefficient.

[0055] An interpretable output generation module is used to summarize the load changes output by the control variable attribution engine, generate readable natural language attribution formulas, and output an analysis report.

[0056] In one implementation, the data fusion and feature construction module, based on multi-site meteorological data fusion and nonlinear extension, includes:

[0057] The meteorological data are weighted according to the assigned meteorological weights to obtain comprehensive meteorological data;

[0058] The comprehensive meteorological data is used to generate square-term features and interaction-term features, which are then used as advanced meteorological features. The square-term features refer to the product of any feature in the comprehensive meteorological data with itself, and the interaction-term features refer to the product of any two different features in the comprehensive meteorological data.

[0059] In one embodiment, the load mechanism separation module includes:

[0060] The frequency domain transformation unit is used to transform the historical load sequence from the time domain to the frequency domain to obtain the spectral coefficients and frequency domain values ​​corresponding to the load data;

[0061] The filtering unit is used to construct a bandpass filter based on a preset target period, extract the load period component from the spectral coefficients, and convert it into a load period component in the time domain.

[0062] The trend calculation unit is used to calculate the load trend component based on the load data and the load cycle component.

[0063] In one implementation, training the regression model includes:

[0064] A linear regression model is used, with the regional meteorological feature set as input and the load trend component as the training target, to fit the model. The complete load data is represented as the sum of the output of the linear regression model, the load cycle component, and a residual component (linear fitting error).

[0065] In one implementation, the single feature change controlled by the control variable attribution engine includes one of the following: date type change, individual meteorological factor change, and higher-order nonlinear feature change.

[0066] In one implementation, the analysis report is in a structured format.

[0067] like Figure 1 As shown, the present invention also provides an interpretable attribution analysis method for power load changes, including the following steps.

[0068] S1: Based on the fusion and nonlinear extension of meteorological data from multiple stations, a regional high-dimensional meteorological feature set is constructed.

[0069] S10: Acquire meteorological and load data.

[0070] Acquire correlated meteorological data from multiple stations in the region and historical load data for that region. The correlated meteorological data is denoted as... Historical load data is recorded as .

[0071] Meteorological data at each station It includes a variety of meteorological features, including but not limited to temperature. ,humidity Irradiation and wind speed In one implementation, meteorological data from each station... It can be represented as , The time column is set at 15-minute intervals. Load data are also presented at 15-minute intervals.

[0072] S11: Using data from multiple meteorological stations in a certain area, weighted and fused according to preset weights to generate comprehensive meteorological data representing the area.

[0073] The weighted average fusion includes:

[0074] Based on the assigned meteorological weights Meteorological data are weighted to obtain comprehensive meteorological data. .

[0075] (1)

[0076] in, , Weight each weather station.

[0077] S12: Expand the fused comprehensive meteorological features to generate a high-dimensional meteorological feature set including high-level meteorological features containing square terms and interaction terms.

[0078] The characteristic of the squared term is:

[0079] (2)

[0080] This represents the i-th feature of the integrated meteorological data.

[0081] The characteristics of the interaction item are:

[0082] (3)

[0083] and Let i and j represent the i-th and j-th features of the integrated meteorological data, respectively.

[0084] S2: Decouple the historical load sequence into load trend components and load cycle components.

[0085] Fourier transform was used to perform frequency domain analysis on the historical load sequence, and the load cycle component of a specific period (T=7 days) was extracted by bandpass filtering. Then, the original load sequence was subtracted from the cycle component to obtain the load trend component.

[0086] In one embodiment, the frequency domain analysis employs Fourier signal decomposition.

[0087] make for This represents load data containing time and load columns; the load data can also be represented in 1*N vector form as follows: Then, using equation (4), a Fast Fourier Transform (FFT) is performed on the load data to transform it from the time domain to the frequency domain, thereby obtaining its spectral coefficients. .

[0088] (4)

[0089] in, This represents the nth value of the load data Y, where i represents the imaginary unit. Then, the corresponding frequency domain value is calculated using equation (5).

[0090] (5)

[0091] If the target period is set to T, then the corresponding target frequency is: Its period T is 7 days, and the filtering tolerance is... Construct the frequency domain coefficients after filtering .

[0092] (6)

[0093] in, This indicates that the absolute value of the frequency is taken, considering the symmetry between positive and negative frequencies. The time-domain periodic components are then reconstructed using an inverse FFT.

[0094] (7)

[0095] income That is, the component with period T=7 in the original sequence, whose real part is taken as the extracted periodic component. The load trend is as follows: .

[0096] (8)

[0097] like Figure 2 As shown, the red line represents the historical load sequence, the green line represents the load trend component, and the blue line represents the load cycle component.

[0098] S3: Train the regression model, map the regional high-dimensional meteorological feature set to the load trend component, and obtain the regression coefficients of each feature.

[0099] In one implementation, the regression model is a linear regression model, with a high-dimensional meteorological feature set as input and the load trend component as the target.

[0100] Using linear regression model The load trend was fitted with meteorological characteristics.

[0101] (9)

[0102] (10)

[0103] Where Q represents the number of features fitted; This is a constant, representing the base load; This represents the q-th feature of the fit; Let q represent the fitting coefficient of the q-th feature.

[0104] Taking the periodic component into account, the fitted load formula can be obtained as follows:

[0105] (11)

[0106] Taking into account the fitting error, the actual load data are:

[0107] (12)

[0108] in, This represents the fitting error.

[0109] In one embodiment, a comparison graph of the load curve fitting curve at a certain moment is shown below. Figure 3 As shown, the model trained by this invention can fit the model almost perfectly, with an average absolute error of 4752.98 and an average absolute percentage error of 4.6%.

[0110] S4: Input the feature data of the analysis day and the comparison day into the trained regression model, and calculate the independent impact of the single feature change on the load trend based on the regression coefficient.

[0111] The controlled variable method was used to calculate the load changes caused by changes in date type, individual meteorological factors, advanced meteorological characteristic factors, and other factors. These other factors refer to the errors generated by the simultaneous linear fitting of multiple input features (original features, squared terms, and interaction terms).

[0112] In one embodiment, it is assumed that the load data at a certain moment on the analysis day is... The characteristics corresponding to a certain moment of the day are analyzed as follows: ; Compare the load data at a certain time on the day to The characteristics corresponding to a certain moment on the day are as follows: ;but:

[0113] The analysis of daily load changes caused by date variations is as follows:

[0114] (13)

[0115] in, and These are the analysis date and the comparison date, respectively.

[0116] Analysis of daily load due to temperature ( The load change caused by the change is:

[0117] (14)

[0118] in, This indicates that only the temperature changes; similarly, the load change caused by each characteristic change can be obtained.

[0119] This embodiment uses temperature and humidity as meteorological features as examples; other features can be processed similarly.

[0120] S5: Summarize the independent impact of each feature on load changes, generate readable natural language attribution formulas, and output a structured analysis report.

[0121] In one embodiment, such as Figure 4 As shown, the comparison curve between the analysis date April 2, 2025 and the analysis date March 20, 2025 is displayed. The load natural language attribution formula for 12:00 on the analysis date April 2, 2025 is obtained.

[0122] Load = Wednesday (225.55) - 8238.78 * Temperature + 146.76 * Temperature² + 37.50 * Temperature x Humidity + 1.71 * Temperature x Irradiance + 95.73 * Temperature x Wind Speed ​​+ 627.90 * Humidity - 6.25 * Humidity² - 0.37 * Humidity x Irradiance - 68.35 * Humidity x Wind Speed ​​+ 8.74 * Irradiance - 0.01 * Irradiance² - 4.55 * Irradiance x Wind Speed ​​+ 3055.71 * Wind Speed ​​+ 52.44 * Wind Speed² + Base Load (128002.20) + Other Factors (-17417.26)

[0123] Based on the attribution formula and the load values ​​and characteristic values ​​of the analysis day and the comparison day, the load change analysis table can be obtained, as shown in Table 1.

[0124]

[0125] To implement the above method, the present invention also provides an electronic device and a computer-readable storage medium.

[0126] The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements some or all of the steps in the above-described method for interpretable attribution analysis of power load changes.

[0127] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements some or all of the steps in the above-described method for interpretable attribution analysis of power load changes.

Claims

1. An interpretable attribution analysis system for power load variations, characterized in that, The system comprises: a data fusion and feature construction module for constructing a regional high-dimensional meteorological feature set based on multi-site meteorological data fusion and nonlinear expansion; a load mechanism separation module for decoupling a historical load sequence into a load trend component and a load cycle component; a quantitative attribution analysis module including a trained regression model for establishing a mapping between the regional high-dimensional meteorological features and the load trend component and storing regression coefficients of each feature; a control variable attribution engine for obtaining high-dimensional meteorological feature data of an analysis day and a comparison day, calling the quantitative attribution analysis module, and calculating a load change amount caused by a single feature change according to the regression coefficients by using a control variable method; an interpretable output generation module for summarizing the load change amount output by the control variable attribution engine, generating a readable natural language attribution formula, and outputting an analysis report.

2. The interpretable attribution analysis system of changes in electric power loads according to claim 1, characterized in that, The data fusion and feature construction module based on multi-site meteorological data fusion and nonlinear expansion comprises: weighting meteorological data according to assigned meteorological weights to obtain comprehensive meteorological data; generating square term features and interaction term features from the comprehensive meteorological data as advanced meteorological features; wherein the square term features refer to the product of any feature in the comprehensive meteorological data and itself, and the interaction term features refer to the product of any two different features in the comprehensive meteorological data.

3. The interpretable attribution analysis system of changes in electric power loads according to claim 1, characterized in that, The load mechanism separation module comprises: a frequency domain conversion unit for converting a historical load sequence from a time domain to a frequency domain to obtain frequency spectrum coefficients and frequency domain values corresponding to load data; a filtering unit for constructing a band-pass filter based on a preset target period, extracting a load cycle component from the frequency spectrum coefficients, and converting the load cycle component into a time-domain load cycle component; a trend calculation unit for calculating a load trend component according to the load data and the load cycle component.

4. The interpretable attribution analysis system of changes in electric power loads according to claim 1, characterized by, The training of the regression model comprises: using a linear regression model, taking the regional meteorological feature set as input, and taking the load trend component as a training target to perform model fitting; complete load data is represented as the sum of the output of the linear regression model, the load cycle component, and a residual component.

5. The interpretable attribution analysis system of changes in electric power loads according to claim 1, characterized by, The single feature change controlled by the control variable attribution engine includes one of the following: date type change, single meteorological factor change, and high-order nonlinear feature change.

6. A method of interpretable attribution of changes in an electric power load, characterized by, The method comprises: constructing a regional high-dimensional meteorological feature set based on multi-site meteorological data fusion and nonlinear expansion; decoupling a historical load sequence into a load trend component and a load cycle component; training a regression model to map the regional high-dimensional meteorological feature set to the load trend component and obtain regression coefficients of each feature; inputting feature data of an analysis day and a comparison day into the trained regression model to calculate an independent influence amount of a single feature change on a load trend according to the regression coefficients; summarizing the independent influence amounts of each feature, generating a readable natural language attribution formula, and outputting an analysis report.

7. The method of claim 6, wherein, The multi-site meteorological data fusion and nonlinear expansion comprises: weighting meteorological data of multiple sites in a region according to meteorological weights assigned thereto to obtain comprehensive meteorological data; The comprehensive weather data is used to generate a square term feature and an interaction term feature as advanced weather features; the square term feature refers to a product of any feature in the comprehensive weather data and itself, and the interaction term feature refers to a product of any two different features in the comprehensive weather data.

8. The method of claim 6, wherein, The mapping of the regional-level high-dimensional weather feature set to the load trend component and the obtaining of regression coefficients of each feature include: The historical load sequence is converted from a time domain to a frequency domain to obtain frequency spectrum coefficients and frequency domain values corresponding to the load data; A band-pass filter is constructed based on a preset target period, a load periodic component is extracted from the frequency spectrum coefficients, and the load periodic component is converted into a time-domain load periodic component; According to the load data and the load periodic component, a load trend component is calculated.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the interpretable attribution analysis method of the power load change according to any one of claims 6 to 8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the interpretable attribution analysis method of the power load change according to any one of claims 6 to 8.

Images