A system and method for photovoltaic power generation and transformer area load prediction

Abstract

The application discloses a kind of for photovoltaic power generation and substation load prediction system and method, the system includes: acquisition module is used to collect distributed photovoltaic power generation power in historical time period, substation load, multi-dimensional environmental parameter and corresponding area's carbon footprint tracking data;Processing module is used to process multi-dimensional environmental parameter and carbon footprint tracking data, extract scene feature vector, through weighted mutual information algorithm and combine carbon influence factor, dynamically filter the core environmental parameter set that adapts to current area;Calculation module is used to calculate the coupling relationship strength between parameters to the core environmental parameter set, combined with regional stability weight obtains joint influence coefficient and lag period;Prediction module is used to construct input sequence by joint influence coefficient and lag period, extract time series feature and coupling feature using prediction model, and output the prediction result of photovoltaic power generation power and substation load.The application improves the accuracy of photovoltaic power generation and substation load prediction.

Description

Technical Field

[0001] This invention belongs to the field of dispatch load data analysis technology, specifically a system and method for photovoltaic power generation and distribution area load forecasting. Background Technology

[0002] Distributed photovoltaic (PV) power generation refers to small-scale PV systems installed on the user side (such as rooftops or industrial parks) to directly supply power or connect to the grid, characterized by intermittency and volatility. Distribution area load forecasting is a quantitative prediction of future electricity demand in a distribution area of ​​the power grid, and the impact of distributed PV output fluctuations on the load needs to be considered.

[0003] In scenarios where a high proportion of distributed photovoltaic (PV) power is integrated into distribution substations, traditional PV power output forecasting and substation load forecasting, with power supply reliability, peak shaving and valley filling, economic cost, and PV consumption as core objectives, and carbon footprint as an evaluation indicator after the fact, are rarely embedded within the forecasting model as input features and constraints. For example, existing technical solutions, such as the invention entitled "A Method and System for Predicting Wind Power," which discloses constructing a multivariate time series forecasting dataset based on historical data of the target to be predicted, training a multivariate time series forecasting model based on the multivariate time series forecasting dataset to obtain a power prediction pre-training model, and using the power prediction pre-training model to predict the power generation of the target within a set future time using the current parameter data of the target to be predicted, have shown excellent performance in time series forecasting tasks in multiple fields. The method proposed in this invention can effectively capture the long-term and short-term dependencies in multivariate time series and improve the accuracy and stability of wind power forecasting by utilizing physical constraints.

[0004] However, existing technical solutions only specify environmental parameters such as wind speed and wind direction, without providing screening criteria and priority ranking mechanisms for other parameters such as temperature, humidity, and air pressure. In practical applications, the key environmental parameters affecting different wind farms vary. For example, humidity has a significant impact on coastal wind farms, while air pressure has a prominent impact on plateau wind farms. Existing solutions cannot dynamically screen core environmental parameters that are suitable for specific scenarios. At the same time, they do not consider the coupling relationship between environmental parameters, such as the synergistic effect of wind speed and wind direction. They only calculate the influence coefficient of individual parameters independently, ignoring the impact of parameter interactions on power generation. As a result, the accuracy of photovoltaic power generation and distribution area load forecasting remains poor. Summary of the Invention

[0005] To address the problems existing in the above solutions, the present invention provides a system and method for photovoltaic power generation and transformer load forecasting.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A photovoltaic power generation and distribution area load forecasting system includes:

[0008] The data acquisition module is used to collect distributed photovoltaic power generation, transformer load, multi-dimensional environmental parameters, and corresponding regional carbon footprint tracking data within a historical time period.

[0009] The processing module is used to process multi-dimensional environmental parameters and carbon footprint tracking data, extract scene feature vectors, and dynamically select the core environmental parameter set that is suitable for the current region through a weighted mutual information algorithm combined with carbon impact factors. Among them, when collecting carbon footprint tracking data, the regional climate type label, carbon emission factor and regional stability weight are obtained through the regional energy management platform interface.

[0010] Extract regional climate type labels from carbon footprint tracking data and convert them into unique thermal encoded vectors;

[0011] The carbon emission factor α is converted to a value in the [0,1] interval using min-max normalization.

[0012] The regional stability weights are calculated based on the variance and smoothing constant of the monthly load data of the transformer area.

[0013] The encoded one-hot encoded vector, the normalized carbon emission factor, and the regional stability weight are concatenated into a scene feature vector.

[0014] Based on the scene feature vector, the mutual information value between each environmental parameter and the target variable is calculated; the normalized carbon emission factor is set as the carbon impact factor, and the mutual information value between the carbon impact factor and each environmental parameter is calculated to obtain the weighted mutual information value of the corresponding environmental parameter.

[0015] All environmental parameters are sorted in descending order of weighted mutual information values ​​to obtain a sorted parameter list; the calculation module is used to calculate the coupling strength between parameters in the core environmental parameter set, and obtain the joint impact coefficient and lag time by combining the regional stability weight;

[0016] The prediction module is used to construct the input sequence by combining the influence coefficient and the lag period, extract the time series features and coupling features using the prediction model, and output the prediction results of photovoltaic power generation and distribution area load.

[0017] Furthermore, the calculation expression for the regional stability weight w is as follows: ;in, The variance of the monthly load data for the transformer substation area; This is the smoothing constant.

[0018] Furthermore, the expression for calculating the weighted mutual information value is: ;in, The weighted mutual information value; Mutual information value; Y is the environmental parameter; Y is the target variable. The normalized carbon emission factor.

[0019] Furthermore, based on the climate type label in the scene feature vector, the number of core parameters N is determined; the first N parameters are selected from the parameter list to form the core environmental parameter set.

[0020] Furthermore, in the calculation module, when calculating the coupling strength between parameters, all non-empty subsets are generated for the core environment parameter set, and all generated non-empty subsets are marked as parameter combinations in sequence;

[0021] The standardized set of core environmental parameters and the sequence of target variables are discretized into K intervals to obtain a discrete sequence;

[0022] Calculate the multivariate mutual information of discrete sequences; multivariate mutual information corresponds to the strength of the coupling relationship between parameters.

[0023] Furthermore, based on the scene feature vector, the coupling strength is transformed into a joint influence coefficient W, and the relevant expression is as follows:

[0024] ;in, For multivariate mutual information of discrete sequences; The carbon emission factor is normalized; w is the regional stability weight. Y is the parameter combination; Y is the target variable.

[0025] Furthermore, regarding parameter combinations The sequence is shifted along the positive time axis by b time points, where b is the offset, to obtain the shifted sequence. ;

[0026] For each offset b, the corresponding multivariate mutual information is calculated. ;

[0027] Utilizing multivariate mutual information of all offsets To determine the lag period, the relevant expression is: ;in, This represents the offset corresponding to the maximum coupling strength, i.e., the lag time of the parameter combination affecting the target variable.

[0028] Furthermore, in the prediction module, a single input sequence is processed by an LSTM sub-model to output temporal features; a target subset sequence is processed by an attention mechanism to output coupled features; and the temporal features and coupled features are weighted and summed to obtain fused features.

[0029] Furthermore, when converting the fusion characteristics into predicted values ​​of photovoltaic power generation and substation load, the fusion characteristics are input into the fully connected layer, and dual predicted values ​​are output:

[0030] ;in, These are the predicted photovoltaic power generation and the predicted transformer load, respectively; Dense is a fully connected layer that maps features to predicted values. This is a feature of fusion.

[0031] A method for photovoltaic power generation and transformer area load forecasting includes:

[0032] Collect historical data on distributed photovoltaic power generation, transformer load, multi-dimensional environmental parameters, and carbon footprint tracking data for the corresponding regions.

[0033] The system processes multi-dimensional environmental parameters and carbon footprint tracking data, extracts scene feature vectors, and dynamically selects a set of core environmental parameters suitable for the current region through a weighted mutual information algorithm combined with carbon impact factors.

[0034] The coupling strength between parameters in the core environmental parameter set is calculated, and the joint impact coefficient and lag time are obtained by combining the regional stability weight.

[0035] The input sequence is constructed by combining the influence coefficient and the lag period. The time series features and coupling features are extracted by the prediction model, and the prediction results of photovoltaic power generation and transformer area load are output.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] This invention addresses the problem of existing technologies being unable to adapt to specific scenarios by preprocessing and extracting features from carbon footprint data, and dynamically screening the core environmental parameter set using a weighted mutual information algorithm combined with carbon impact factors. It also incorporates carbon impact factors to enhance the relevance of the parameters to the carbon targets, providing efficient and suitable input features for subsequent prediction models. Furthermore, it quantifies parameter coupling strength using multivariate mutual information, generates joint impact coefficients by combining regional stability weights, and determines lag periods through time-shift analysis, overcoming the problem of existing technologies neglecting parameter interactions and lags. This provides more accurate feature inputs for distributed photovoltaic and transformer load forecasting. By considering both lag and coupling, the accuracy of photovoltaic power generation and transformer load forecasting is improved, providing reliable support for grid optimization and dispatching. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a system block diagram of the present invention.

[0040] Figure 2 This is a flowchart illustrating the process of obtaining the joint influence coefficient in this invention.

[0041] Figure 3 This is a flowchart illustrating the process of outputting two predicted values ​​in this invention.

[0042] Figure 4 This is a flowchart of the method of the present invention. Detailed Implementation

[0043] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0044] Example 1, such as Figure 1 As shown, a photovoltaic power generation and distribution area load forecasting system includes a data acquisition module, a processing module, a calculation module, and a forecasting module.

[0045] The data acquisition module collects distributed photovoltaic power generation, transformer load, multi-dimensional environmental parameters, and corresponding regional carbon footprint tracking data over historical time periods. Specific steps include:

[0046] Based on the power metering boundary of the distribution area, such as the power supply range of a 10kV distribution area, record the latitude and longitude coordinates of the distributed photovoltaic units and the list of residential and commercial users covered by the distribution area load to ensure that the data covers all objects in the prediction scenario; among them, distributed photovoltaic units include, for example, residential and rooftop photovoltaics;

[0047] Historical time period, specifically the past 30 days;

[0048] When collecting photovoltaic power data, a data acquisition module is deployed on the RS485 port of the distributed photovoltaic inverter to read the active power and reactive power output by the inverter in real time and transmit them to the local data server.

[0049] When collecting data on the load of a distribution area, the total load and load data of each branch circuit in the distribution area are collected through the LoRa interface of the smart meter in the distribution area. After being synchronized to the power marketing system, the data is forwarded to the local server through the API interface.

[0050] When collecting multi-dimensional environmental parameters, a meteorological sensor group is deployed at the center of the photovoltaic panel array, including a light intensity sensor, a temperature sensor, a humidity sensor, and a barometric pressure sensor, which collect the corresponding light intensity, temperature, humidity, and barometric pressure data respectively; the sensor data is wirelessly transmitted to the server via NB-IoT.

[0051] When collecting carbon footprint tracking data, regional climate type labels, carbon emission factors, and regional stability weights are obtained through the regional energy management platform interface. The regional climate type labels, such as "coastal humid" or "high-altitude cold," are automatically generated based on the region's meteorological statistics for the past three years. Specifically:

[0052] The coastal areas are humid, specifically defined as an annual average humidity >80% and an air pressure <100kPa;

[0053] The plateau is cold, specifically corresponding to an annual average air pressure of <90kPa and a temperature of <15℃;

[0054] The carbon emission factor α, expressed in kgCO2 / kWh, is obtained from the regional power grid carbon emission report. It represents the amount of carbon emissions reduced by generating 1 kWh of photovoltaic power.

[0055] The calculation expression for the regional stability weight w is as follows: ;in, The variance of the monthly load data for the transformer substation reflects the degree of fluctuation. The smoothing constant is set to 0.001 to avoid a denominator of 0; the regional stability weight w ranges from (0,1], and the larger the value, the more stable the regional load.

[0056] In this embodiment of the invention, by performing multi-dimensional data acquisition, diverse data support can be provided for subsequent dynamic parameter screening and coupling relationship analysis, thereby improving the accuracy and reliability of data analysis.

[0057] The processing module processes carbon footprint tracking data, extracts scene feature vectors, and dynamically selects a set of core environmental parameters suitable for the current region using a weighted mutual information algorithm combined with carbon impact factors. Specific steps include:

[0058] Extract regional climate type labels from carbon footprint tracking data and convert them into unique thermal encoding vectors L. For example, coastal humid L1 corresponds to [1,0,0], and plateau cold L2 corresponds to [0,1,0].

[0059] The carbon emission factor α is converted to a value in the [0,1] interval using min-max normalization, and the relevant expression is as follows: Where α is the original carbon emission factor; It is the minimum carbon emission factor for similar regions; This represents the maximum carbon emission factor for similar regions.

[0060] The encoded unique heat encoding vector L and the normalized carbon emission factor are used to... The regional stability weights w are concatenated to form the scene feature vector. ;

[0061] It should be noted that by converting unstructured carbon footprint data into structured numerical vectors, it can be directly used for subsequent weighted mutual information calculations, enabling correlation analysis between carbon footprint and environmental parameters.

[0062] When calculating the mutual information value between each environmental parameter and the target variable based on the scene feature vector, the target variable Y is determined according to the prediction task. , These are distributed photovoltaic power and transformer area load, respectively; the forecasting task includes photovoltaic power forecasting and transformer area load forecasting.

[0063] For each environmental parameter Examples include temperature, humidity, and light intensity, where i represents the index of different environmental parameters, and the mutual information between these parameters and the target variable Y is calculated. The formula is:

[0064] Where p(x,y) is The joint probability distribution of Y; p(x) is an environmental parameter. The marginal probability distribution of the target variable Y is given by p(y); x represents the specific values ​​of the environmental parameters; and y represents the specific values ​​of the target variable.

[0065] By concatenating and combining all the calculated mutual information, a mutual information sequence is obtained. ;

[0066] It should be noted that by calculating the mutual information between each environmental parameter and the target variable Y, the influence of a single environmental parameter on the target variable can be quantified, providing a basic indicator for subsequent weighted screening.

[0067] Normalized carbon emission factors Set as the carbon impact factor for each environmental parameter Calculate the weighted mutual information value The formula is:

[0068] ;

[0069] All environmental parameters are weighted according to mutual information values Sort the parameters in descending order to obtain the sorted list.

[0070] It is important to explain that adjusting environmental parameters by combining the emission reduction characteristics of carbon footprint makes the screening results more aligned with regional carbon target requirements.

[0071] When dynamically filtering and adapting the core environmental parameter set to the current region, the number of core parameters N is determined based on the climate type label L in the scene feature vector V. For example, coastal humid L1 corresponds to N=5, and plateau cold L2 corresponds to N=6. The specific value of the number of core parameters can also be customized by professionals in this field according to the application requirements of the actual application scenario.

[0072] Select the first N parameters from the parameter list to form the core environment parameter set. .

[0073] It is important to note that dynamically adjusting the number of core parameters based on regional climate types, and ensuring that the screening results are adapted to specific scenarios, can reduce the impact of redundant parameters on prediction performance.

[0074] In this embodiment of the invention, by preprocessing and extracting features from carbon footprint data, and by using a weighted mutual information algorithm combined with carbon impact factors to dynamically screen the core environmental parameter set, the problem that existing technologies cannot adapt to specific scenarios is solved. At the same time, the integration of carbon impact factors improves the relevance of the carbon targets in parameter screening, which can provide efficient and suitable input features for subsequent prediction models.

[0075] The calculation module calculates the coupling strength between parameters in the core environmental parameter set, and combines this with regional stability weights to obtain the joint impact coefficient and lag time. Specific steps include:

[0076] Each parameter in the core environment parameter set is z-score standardized, which is a conventional technical solution. The specific implementation steps are not detailed here.

[0077] like Figure 2 As shown, when calculating the coupling strength between parameters, all non-empty subsets are generated for the core environment parameter set. These non-empty subsets are then sequentially labeled as parameter combinations, such as pairwise combinations. ternary combination ;

[0078] The standardized set of core environmental parameters and the sequence of target variables are discretized into K intervals to obtain a discrete sequence; K is a positive integer, with a default value of 10.

[0079] Calculating multivariate mutual information of discrete sequences :

[0080] ;in, A non-empty subset generated for the core environment parameter set; The joint entropy of the parameter combination, , This represents the probability distribution of the discretized parameter combinations; The entropy of the target variable, ; The joint entropy of the parameter combination and the target variable. Multivariate mutual information corresponds to the strength of coupling between parameters;

[0081] It should be noted that by calculating the multivariate mutual information of discrete sequences, the degree of synergistic influence between parameters can be quantified, making up for the deficiency of single-parameter analysis in ignoring interactions, such as the synergistic effect of light and temperature on photovoltaic power.

[0082] Based on scene feature vectors The coupling strength is transformed into the joint influence coefficient W, and the relevant expression is as follows:

[0083] ;

[0084] It is worth noting that by incorporating carbon footprint characteristics into the coupled impact assessment, the joint impact coefficient can reflect both the interaction of parameters and align with regional carbon emission reduction targets and actual operational stability.

[0085] In addition, the joint impact coefficient can also be calculated in another way, specifically:

[0086] By utilizing a normalized carbon emission factor, the contribution to carbon emission reduction is incorporated into the evaluation dimension of the joint impact coefficient. This allows for parameter combination selection while simultaneously considering the accuracy of power prediction and its relevance to carbon targets, achieving end-to-end connectivity of carbon footprint data. Based on the original multivariate mutual information, a carbon impact factor weighting term is introduced to obtain the dual-objective coupling strength. ;in, To enhance the dual-target coupling strength for merging carbon targets; Non-empty subsets generated for the core environment parameter set The carbon sensitivity coefficient characterizes the degree of influence of the parameter combination on the carbon footprint of the distribution area, and the value range is [0,1]. The combination of photovoltaic strongly correlated parameters such as irradiance and temperature is 0.8~1, the combination of load-related parameters is 0.4~0.6, and the combination of pure meteorological auxiliary parameters is 0.1~0.3.

[0087] A dimensionality penalty term is introduced to penalize and correct subsets with an excessive number of parameters: ;in, The effective coupling strength after dimensionality penalty; Non-empty subsets generated for the core environment parameter set The number of parameters included; The penalty coefficient ranges from [0.05, 0.2]. For high-dimensional scenarios, it is 0.15 to 0.2, and for low-dimensional scenarios, it is 0.05 to 0.1. It is used to control the penalty intensity.

[0088] Based on the climate type unique thermal encoding vector L, a climate scene weight matrix is ​​generated, and scene correction is performed on the effective coupling strength: ;in, It is the coupling strength after climate scenario correction; This is a baseline weight vector for climate types, with dimensions matching the number of climate types. Each element corresponds to the baseline weight of a subset of parameters under that climate. For example, in humid coastal areas, the weight of the parameter combination including humidity and air pressure is 1.2, and the weight of the parameter combination including temperature is 1. In cold plateau areas, the weight of the parameter combination including temperature and radiation is 1.2, and the weight of the parameter combination including humidity is 0.9.

[0089] Constructing three-dimensional comprehensive stability weights Replace the original single regional stability weight w:

[0090] ;in, The original region stability weight, w; The data stability weights for the non-empty subset generated from the core environmental parameter set are obtained by normalizing the standard deviation of the time-series data of the parameter subset, and the calculation formula is as follows: , The time-series standard deviation of the parameter subset. The maximum standard deviation of all parameter subsets; The data integrity weight of the parameter subset is calculated by the proportion of valid data points in the 30-day collection period to the total number of data points, and the value range is [0,1]. These are the allocation coefficients for the three-dimensional weights. The default values ​​are 0.4, 0.3, and 0.3 respectively.

[0091] Introducing a time-delay matching factor, which, together with coupling strength and stability weights, synthesizes a joint influence coefficient: ;in, The time delay matching factor is a subset of parameters, with a value range of [0,1]. The larger the value, the more accurate and stable the time delay matching. For the parameter subset in the optimal lag time The time-series correlation coefficient between the target variable and the time-delay variable reflects the degree of time-delay matching; To determine the optimal lag period, for each parameter subset, the sliding time shift method is used to iterate through different lag orders, calculate the time-series correlation coefficient with the target variable, and take the order corresponding to the maximum value as the optimal lag period. The standard deviation of the optimal lag order of the parameter subset under the sliding time window reflects the time delay stability. A 5-day sliding time window is used to calculate the optimal lag order under each window and calculate its standard deviation. The maximum value of the standard deviation of the lag order of all parameter subsets is used for normalization;

[0092] The final complete expression for the joint influence coefficient is: ;

[0093] It should be noted that the above processing extends the carbon impact factor from single-parameter weighting to multi-parameter combination evaluation, achieving end-to-end effectiveness of carbon footprint data from collection and preprocessing to parameter selection. The parameter selection results are simultaneously adapted to power prediction and low-carbon requirements. Climate type is upgraded from simply controlling the number of parameters to differentiated weighting correction, strengthening the weight of dominant parameter combinations in different climate zones and suppressing the weight of irrelevant parameters. The dimensionality penalty term addresses the inherent bias of multivariate mutual information towards multiple parameters, avoiding redundant parameter stacking and compressing the input dimension while ensuring coupling strength. Through the synergy of three-dimensional stability weights and time-lag matching factors, the effectiveness of coefficients is constrained from two dimensions: data quality and time series alignment, which can reduce the interference of abnormal data and time series misalignment on prediction.

[0094] For parameter combinations The sequence is shifted b time points along the positive time axis, where b is the offset ranging from [0, T / 2], and T is the total number of time points in the historical time period, resulting in the shifted sequence. ;

[0095] For each offset b, the corresponding multivariate mutual information is calculated. ;

[0096] Utilizing multivariate mutual information of all offsets To determine the lag period, the relevant expression is: ;in, This represents the offset corresponding to the maximum coupling strength, i.e., the lag time of the parameter combination's effect on the target variable, for example... This indicates the effect of parameter combinations on the target variable after two time points.

[0097] It should be noted that by considering the time lag of the effect of parameter coupling, the time alignment accuracy of the input sequence of the prediction model can be ensured, thereby improving the prediction accuracy.

[0098] In this embodiment of the invention, the coupling strength of parameters is quantified by multivariate mutual information, and a joint influence coefficient is generated by combining regional stability weights. The lag period is determined by time shift analysis, which solves the problem of neglecting parameter interaction and lag in the prior art. This can provide more accurate feature input for distributed photovoltaic and transformer area load forecasting.

[0099] The prediction module constructs the input sequence by combining the influence coefficient and the lag period, extracts time-series and coupling features using the prediction model, and outputs the prediction results for photovoltaic power generation and transformer load. The specific steps include:

[0100] When constructing the input sequence using the joint impact coefficient and lag period, the joint impact coefficient is filtered, and parameter combinations with joint impact coefficients greater than or equal to the impact threshold are retained and marked as target subsets. The default value of the impact threshold is 0.6 to ensure that the input sequence contains key coupling information.

[0101] like Figure 3 As shown, for each filtered target subset, the current time is taken as the end point of the time window, shifted backward by τ time points, and the data within the window is taken as the input sequence. The relevant expression is: Where S is the input sequence of the target subset; The time window length, for example, 12 time points; Let t be the starting point of the window, taking into account lag; t is the current time.

[0102] It should be noted that the constructed input sequence can not only capture the lag of parameter coupling, but also screen key combinations through joint influence coefficients to reduce redundant information, which can effectively improve prediction efficiency and accuracy.

[0103] When extracting temporal and coupled features using a prediction model, a multi-sub-model architecture is used to extract both types of features, which are then fused to enhance expressive power; specifically:

[0104] A single input sequence is processed by an LSTM sub-model to output temporal features. These features reflect the time trend of the parameters, and the relevant expressions are as follows:

[0105] ;in, Let i be the temporal characteristics of the i-th environmental parameter; The input sequence is LSTM, which is a Long Short-Term Memory network used to capture temporal dependencies.

[0106] The LSTM sub-model uses a two-layer bidirectional LSTM structure with 64 hidden units per layer, tanh activation function, and a fully connected output layer. Its dimensions match those of the subsequent fused features.

[0107] ; where Bidirectional is a bidirectional wrapper used to convert a unidirectional recurrent neural network into a bidirectional recurrent neural network; This is the first LSTM layer, and 64 is the number of hidden units in the LSTM layer. This is the output mode parameter for the LSTM layer. When True, the LSTM layer outputs the hidden state sequence for each time step; when False, it only outputs the hidden state of the last time step. The second LSTM layer is Dense(32), which is a fully connected layer.

[0108] Furthermore, the target subset sequence is processed through an attention mechanism to output coupling features. These coupling features reflect the synergistic effect between parameters, and the relevant expression is as follows:

[0109] ;in, The coupling characteristics of the target subset; The target subset sequence; Attention is the attention mechanism that uses the joint influence coefficient W as the weight to enhance the features of high-contribution combinations;

[0110] Specifically, the temporal features of all environmental parameters are concatenated along the feature dimension to form a target subset sequence. When using scaled dot product attention to dynamically learn the weights of the parameter combination sequence, a linear transformation is performed on the parameter combination sequence to obtain the query Q, key K, and value V matrix. The attention score is calculated and scaled using the query Q, key K, and value V matrix, and finally the attention weights are obtained by applying the softmax function.

[0111] The attention weights are then applied to the value matrix to generate a weighted feature sequence. The scaling dot product attention dynamic learning, linear transformation, and calculation and application of attention weights are all existing conventional technical solutions, and the specific implementation steps will not be elaborated here.

[0112] The weighted feature sequence is input into the fully connected layer to fuse features, and the output is a coupled feature.

[0113] The fused features are obtained by weighted summation of temporal features and coupled features. :

[0114] ;in, The influence coefficient of a single parameter; The joint influence coefficient of the target subset; the influence coefficient of a single parameter and the joint influence coefficient of the target subset can be customized by professionals in this field according to the application requirements and specifications of the actual application scenario, or can be determined by simulation training and testing based on the sample data, according to the maximum or median value corresponding to the simulation training and testing data, and the specific value is not limited.

[0115] It should be noted that by simultaneously capturing the time trend of individual parameters and the synergistic effect between parameters, the feature expression is more comprehensive, which can lay the foundation for accurate prediction.

[0116] When converting the fusion characteristics into predicted values ​​of photovoltaic power generation and substation load, the fusion characteristics are input into the fully connected layer, and dual predicted values ​​are output:

[0117] ;in, These are the predicted photovoltaic power generation and the predicted transformer load, respectively; Dense is a fully connected layer that maps features to predicted values.

[0118] Perform constraint analysis on the two predicted values ​​to ensure they are within a reasonable range:

[0119] ; ;in, This represents the maximum output power of the photovoltaic system. This represents the maximum load in the transformer area.

[0120] The final predicted value is output based on the results of constraint analysis, supporting the dual-dimensional requirements of power grid dispatch.

[0121] It should be noted that the above steps enable joint forecasting of photovoltaic power generation and distribution area load, thus meeting the grid dispatching requirements for coordinated decision-making on renewable energy output and load demand.

[0122] In this embodiment of the invention, by taking into account both hysteresis and coupling, the accuracy of photovoltaic power generation and distribution area load forecasting is improved, which can provide reliable support for grid optimization and dispatching.

[0123] Example 2, as Figure 4 As shown, a method for photovoltaic power generation and distribution area load forecasting includes:

[0124] Collect historical data on distributed photovoltaic power generation, transformer load, multi-dimensional environmental parameters, and carbon footprint tracking data for the corresponding regions.

[0125] The system processes multi-dimensional environmental parameters and carbon footprint tracking data, extracts scene feature vectors, and dynamically selects a set of core environmental parameters suitable for the current region through a weighted mutual information algorithm combined with carbon impact factors.

[0126] The coupling strength between parameters in the core environmental parameter set is calculated, and the joint impact coefficient and lag time are obtained by combining the regional stability weight.

[0127] The input sequence is constructed by combining the influence coefficient and the lag period. The time series features and coupling features are extracted by the prediction model, and the prediction results of photovoltaic power generation and transformer area load are output.

[0128] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0129] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0130] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0131] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A photovoltaic power generation and distribution area load forecasting system, characterized in that, include: The data acquisition module is used to collect distributed photovoltaic power generation, transformer load, multi-dimensional environmental parameters, and corresponding regional carbon footprint tracking data within a historical time period. The processing module is used to process multi-dimensional environmental parameters and carbon footprint tracking data, extract scene feature vectors, and dynamically select the core environmental parameter set that is suitable for the current region through a weighted mutual information algorithm combined with carbon impact factors. Among them, when collecting carbon footprint tracking data, the regional climate type label, carbon emission factor and regional stability weight are obtained through the regional energy management platform interface. Extract regional climate type labels from carbon footprint tracking data and convert them into unique thermal encoded vectors; The carbon emission factor α is converted to a value in the [0,1] interval using min-max normalization. The regional stability weights are calculated based on the variance and smoothing constant of the monthly load data of the transformer area. The encoded one-hot encoded vector, the normalized carbon emission factor, and the regional stability weight are concatenated into a scene feature vector. Based on the scene feature vector, the mutual information value between each environmental parameter and the target variable is calculated; the normalized carbon emission factor is set as the carbon impact factor, and the mutual information value between the carbon impact factor and each environmental parameter is calculated to obtain the weighted mutual information value of the corresponding environmental parameter. Sort all environmental parameters in descending order of weighted mutual information values ​​to obtain a sorted parameter list; The calculation module is used to calculate the coupling strength between parameters in the core environmental parameter set, and combine it with regional stability weights to obtain the joint impact coefficient and lag time. The prediction module is used to construct the input sequence by combining the influence coefficient and the lag period, extract the time series features and coupling features using the prediction model, and output the prediction results of photovoltaic power generation and distribution area load.

2. The photovoltaic power generation and distribution area load forecasting system according to claim 1, characterized in that, The calculation expression for the regional stability weight w is as follows: ;in, The variance of the monthly load data for the transformer substation area; This is the smoothing constant.

3. A photovoltaic power generation and distribution area load forecasting system according to claim 2, characterized in that, The expression for calculating the weighted mutual information value is: ;in, The weighted mutual information value; Mutual information value; Y is the environmental parameter; Y is the target variable. The normalized carbon emission factor.

4. A photovoltaic power generation and distribution area load forecasting system according to claim 3, characterized in that, Based on the climate type label in the scene feature vector, determine the number N of core parameters; select the first N parameters from the parameter list to form the core environmental parameter set.

5. A photovoltaic power generation and distribution area load forecasting system according to claim 4, characterized in that, In the calculation module, when calculating the coupling strength between parameters, all non-empty subsets are generated for the core environment parameter set, and all generated non-empty subsets are marked as parameter combinations in sequence. The standardized set of core environmental parameters and the sequence of target variables are discretized into K intervals to obtain a discrete sequence; Calculate the multivariate mutual information of discrete sequences; multivariate mutual information corresponds to the strength of the coupling relationship between parameters.

6. A photovoltaic power generation and distribution area load forecasting system according to claim 5, characterized in that, Based on the scene feature vector, the coupling strength is transformed into a joint influence coefficient W, and the relevant expression is as follows: ;in, For multivariate mutual information of discrete sequences; The carbon emission factor is normalized; w is the regional stability weight. Y is the parameter combination; Y is the target variable.

7. A photovoltaic power generation and distribution area load forecasting system according to claim 6, characterized in that, For parameter combinations The sequence is shifted along the positive time axis by b time points, where b is the offset, to obtain the shifted sequence. ; For each offset b, the corresponding multivariate mutual information is calculated. ; Utilizing multivariate mutual information of all offsets To determine the lag period, the relevant expression is: ;in, This represents the offset corresponding to the maximum coupling strength, i.e., the lag time of the parameter combination affecting the target variable.

8. A photovoltaic power generation and distribution area load forecasting system according to claim 7, characterized in that, In the prediction module, a single input sequence is processed through an LSTM sub-model to output time-series features; The target subset sequence is processed by an attention mechanism to output coupled features; the temporal features and coupled features are weighted and summed to obtain fused features.

9. A photovoltaic power generation and distribution area load forecasting system according to claim 8, characterized in that, When converting the fusion characteristics into predicted values ​​of photovoltaic power generation and substation load, the fusion characteristics are input into the fully connected layer, and dual predicted values ​​are output: ;in, These are the predicted photovoltaic power generation and the predicted transformer load, respectively; Dense is a fully connected layer that maps features to predicted values. This is a feature of fusion.

10. A method for photovoltaic power generation and transformer load forecasting, applied to a photovoltaic power generation and transformer load forecasting system as described in any one of claims 1-9, characterized in that, include: Collect historical data on distributed photovoltaic power generation, transformer load, multi-dimensional environmental parameters, and carbon footprint tracking data for the corresponding regions. The system processes multi-dimensional environmental parameters and carbon footprint tracking data, extracts scene feature vectors, and dynamically selects a set of core environmental parameters suitable for the current region through a weighted mutual information algorithm combined with carbon impact factors. The coupling strength between parameters in the core environmental parameter set is calculated, and the joint impact coefficient and lag time are obtained by combining the regional stability weight. The input sequence is constructed by combining the influence coefficient and the lag period. The time series features and coupling features are extracted by the prediction model, and the prediction results of photovoltaic power generation and transformer area load are output.

Images