Regional charge source prediction method and device based on distributed multi-source data

Abstract

The invention provides a regional charge source prediction method based on distributed multi-source data. The regional charge source prediction method comprises the steps of 1, acquiring historical parameter data of distributed power equipment in a region and historical total charge source data in the region; 2, preprocessing the historical parameter data of each piece of distributed power equipment; 3, constructing a regional load source prediction model, and training the regional load source prediction model based on the preprocessed historical parameter data and historical total load source data in the region; and step 4, inputting historical parameter data of each distributed power device in the region of the prediction target into the region charge source prediction model, outputting a region total charge source prediction value in the future preset time, and realizing prediction of the region total charge source in the future preset time. According to the method, integrated prediction of the load demand and the power supply output in the region is realized, and the internal coupling relationship of the load source is fully learned, so that a power management department can more intuitively and conveniently perform power resource scheduling work.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power systems, and in particular to a regional load-source prediction method based on distributed multi-source data. BACKGROUND

[0002] With the acceleration of urbanization, the structure and operation mode of regional power grids are undergoing profound changes. New energy power sources, new energy storage devices, electric vehicles, and diversified intelligent loads, etc. new type of distributed power facilities have been widely accessed and applied on the basis of the original power grid system. This not only improves the flexibility and intelligence level of the power system, but also promotes the consumption of renewable energy and green and low-carbon development. However, the large-scale access of new type of distributed power facilities has generated a large amount of multi-source heterogeneous power data, and at the same time has led to stronger volatility and uncertainty of regional power load demand and power output. Under this background, accurately predicting regional power load demand and power output is a key direction for the development of future power systems, and plays an important role in efficient scheduling of power resources, ensuring safe and stable operation of power grids, and optimizing energy allocation.

[0003] Most of the existing regional load-source prediction methods regard power load demand and power output as two independent prediction tasks, and in terms of data utilization, the existing technology usually only relies on regional-level overall data such as historical load curves, historical power output data, and meteorological factors (such as temperature, wind speed, humidity). Although the existing technology can capture macro trends to some extent, it has obvious limitations. First, there is a complex internal coupling relationship between load demand and power output, and the separation of the two will make it difficult for the model to fully learn and characterize this coupling characteristic; second, the regional-level overall data has a large granularity, which will make the prediction model unable to learn the influence of multi-source distributed access devices on the dynamic changes of regional overall load source. Therefore, the generalization ability and accuracy of the existing regional load-source prediction method are insufficient, and it is difficult to meet the actual needs of the high-quality development of smart grids.

[0004] In view of this, in order to overcome the above-mentioned deficiencies, it becomes a problem to be solved in the field to provide a regional load-source prediction method based on distributed multi-source data. SUMMARY

[0005] The purpose of the present application is to provide a regional load-source prediction method based on distributed multi-source data, by constructing a regional load-source prediction model and training the prediction model based on each distributed power device in the region and historical total load source data in the region, to realize comprehensive and accurate prediction of load demand and power output in the region, to help the power management department to realize efficient scheduling of power resources, to ensure safe and stable operation of power grids, and to optimize energy allocation.

[0006] To solve the above technical problems, the technical scheme of the present application is:

[0007] In a first aspect, the present application provides a regional load source prediction method based on distributed multi-source data. It includes the following steps:

[0008] Step 1: Collecting original data, obtaining historical parameter data of each distributed power equipment in the region and historical total load source data in the region;

[0009] Step 2: Preprocessing the historical parameter data of each distributed power equipment;

[0010] Step 3: Constructing a regional load source prediction model, and training the regional load source prediction model based on the preprocessed historical parameter data and the historical total load source data in the region, the regional load source prediction model including a distributed equipment end and a regional total network end, which is implemented as follows:

[0011] Step 3.1: Input the preprocessed historical parameter data into the distributed equipment end, and perform preliminary fusion through a lightweight coding fusion mechanism to output the load source fusion features of the distributed power equipment;

[0012] Step 3.2: Input the load source fusion features of the distributed power equipment and the historical total load source data in the region into the regional total network end through the attention mechanism for deep fusion and decoding prediction to obtain the total load source prediction result in the region;

[0013] Step 4: Input the historical parameter data of each distributed power equipment in the region of the prediction target into the regional load source prediction model, and output the regional total load source prediction value in the future preset time to realize the prediction of the regional total load source in the future preset time.

[0014] Preferably, the distributed power equipment in step 1 includes a photovoltaic array group, a wind energy group, a energy family user unit, an electric vehicle charging pile and an energy storage device group, and the historical load source data in the region includes total load demand data in the region and total power output data in the region.

[0015] Preferably, the preprocessing of the historical parameter data of each distributed power equipment collected in step 2 includes data cleaning and normalization processing.

[0016] Preferably, step 3.1 includes the following steps:

[0017] Step 3.1.1: Based on the preprocessed historical parameter data of each distributed power equipment, constructing an input sequence of each distributed power equipment;

[0018] Step 3.1.2: The input sequence is respectively sent to the local encoding fusion module corresponding to the distributed power equipment, and a lightweight encoding fusion mechanism is used for preliminary fusion to obtain the load source fusion feature of each distributed power equipment, which is implemented as follows:

[0019] (a) The input sequence with dimension and length is respectively multiplied by the query parameter matrix, the key parameter matrix and the value parameter matrix with size to obtain the query matrix, the key matrix and the value matrix with size

[0020] (b) The query matrix is weighted aggregated in the time sequence dimension, and the aggregated sequence is normalized to obtain the query sequence with size

[0021] (c) The query sequence is multiplied by the key matrix element by element, and is added along the dimension direction to obtain the global query sequence with size

[0022] (d) The global query sequence is multiplied by the value matrix element by element to obtain the fusion feature matrix with size

[0023] (e) The fusion feature matrix is flattened and converted to the load source fusion feature with dimension 1 and length after the full connection layer.

[0024] Preferably, the lightweight encoding fusion mechanism has the following calculation formula: wherein, are the query matrix, the key matrix and the value matrix of the lightweight fusion mechanism in the local encoding fusion module, are the query parameter matrix, the key parameter matrix and the value parameter matrix of the lightweight fusion mechanism in the local encoding module, is the input sequence, is the query sequence, is the normalization function, represents the time sequence direction aggregation, is the global query sequence, represents the dimension direction addition, is the load source fusion feature, represents the full connection layer, represents the flattening.​​​​

[0025] Preferably, step 3.2 comprises the following steps:

[0026] Step 3.2.1: The source load fusion features output in step 3.1 are spliced in the dimension direction to obtain a spliced matrix with a size of ;

[0027] Step 3.2.2: The spliced matrix is deep fused through self-attention mechanism via serial total network coding fusion modules to obtain regional total network coding fusion features with a size of , and the specific implementation is as follows:

[0028] (a) The spliced matrix with a size of is taken as the input of the first total network coding fusion module, and is multiplied by a query parameter matrix with a size of , a key parameter matrix with a size of , and a value parameter matrix with a size of , respectively, to obtain a query matrix with a size of

[0029] , a key matrix with a size of , and a value matrix with a size of

[0030] ; (c) The normalized attention matrix is multiplied by the value matrix to obtain self-attention output with a size of

[0031] ; (d) The self-attention output with a size of is taken as the input of the next serial total network coding fusion module, and steps (a) to (d) are repeatedly executed for layer-by-layer fusion until the deep fusion of serial total network coding fusion modules is completed to obtain regional total network coding fusion features with a size of

[0032] , which are then output;

[0033] Step 3.2.3: The regional total network coding fusion features and the historical total load data in the region are taken as the input of the decoding prediction module, and are decoded and predicted through a supervised attention mechanism to obtain the total load prediction result in the region, and the specific implementation is as follows: (a) The regional total network coding fusion features are multiplied by a query parameter matrix with a size of to obtain a query matrix with a size of ; the historical total load data in the region are multiplied by a key parameter matrix with a size of and a value parameter matrix with a size of

[0034] , respectively, to obtain a key matrix with a size of and a value matrix with a size of ;(b) By using a self-attention mechanism, the query matrix containing the regional total network coding fusion feature information is decoded and interacted with the key matrix and value matrix containing the historical total load source data information within the region, resulting in a matrix of size [missing information]. The decoding matrix;

[0035] (c) Flatten the decoding matrix as A vector of length is transformed through a fully connected layer into... The predicted length, where the first The first forecast value is the predicted total regional load demand value for a future time period, followed by... Each forecast value represents the predicted total power output of the region over a future period.

[0036] Preferably, step 3 further includes using mean square error to converge the model and optimize the model parameters during the training process of the regional load source prediction model.

[0037] Preferably, in step 3.2.2, the self-attention mechanism in the overall network coding fusion module is calculated using the following formula: in, These are the query matrix, key matrix, and value matrix of the self-attention mechanism in the overall network coding fusion module, respectively. To splice the matrix, These are the query parameter matrix, key parameter matrix, and value parameter matrix of the self-attention mechanism in the overall network coding fusion module, respectively. Table normalization function, This is a feature of regional network coding fusion.

[0038] Preferably, in step 3.2.3, the supervised attention mechanism in the decoding prediction module is calculated using the following formula: in, These are the query matrix, key matrix, and value matrix of the supervised attention mechanism in the decoding and prediction module, respectively. For the regional network coding fusion characteristics, These are the query parameter matrix, key parameter matrix, and value parameter matrix of the supervised attention mechanism in the decoding and prediction module, respectively. This refers to the historical total load source data for the region. This represents the predicted total load source value within the region. Table normalization function.

[0039] In a second aspect, the present invention provides a regional load source prediction device based on distributed multi-source data, comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is called and executed by the processor, it implements the regional load source prediction method based on distributed multi-source data as described in any one of claims 1-9.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) For existing load demand or power output prediction research, most studies regard them as two independent prediction tasks. However, from the perspective of power management departments, a dynamic balance should be maintained between power output and load demand. If the predicted load demand is higher than the power output, preparations should be made to increase power production capacity; conversely, if the predicted power output is higher than the load demand, appropriate reduction of power generation should be considered. Therefore, compared with most studies that treat load demand and power output separately, the present invention starts from the massive parameter data of the underlying distributed equipment through the regional load source prediction model. Through the local coding fusion of the distributed equipment end and the total network coding fusion and decoding prediction of the regional total network end, the integrated prediction of load demand and power output in the region is realized. The inherent coupling relationship between load and source is fully learned, so that the power management department can carry out power resource scheduling work more intuitively and conveniently; (2) The regional load source prediction model proposed in the present invention includes the distributed equipment end and the regional total network end. On the one hand, a lightweight coding fusion mechanism is proposed at the distributed device level. This mechanism compresses the length of the query matrix through aggregation operations and replaces matrix multiplication in the original self-attention mechanism with element-wise multiplication, thereby improving the coding fusion efficiency of the distributed devices and saving computing resources at the distributed device level. On the other hand, a supervised attention mechanism is proposed at the regional total network level. This mechanism introduces historical data of the total load sources in the region as additional auxiliary supervision, which helps to alleviate the problem of regional total network coding fusion feature offset caused by data distortion of some distributed devices, improves the fitting ability of the regional load source prediction model, and further ensures the robustness of the prediction model. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention 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.

[0042] Figure 1 This is a flowchart illustrating the regional load source prediction method based on distributed multi-source data of the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of the regional load source prediction model of the present invention;

[0044] Figure 3 This is a schematic diagram of the local encoding fusion module in the distributed device of the present invention;

[0045] Figure 4 This is a schematic diagram of the central network encoding fusion module in the regional central network terminal of the present invention;

[0046] Figure 5 This is a schematic diagram of the decoding and prediction module in the regional network terminal of the present invention. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings. The terminal technical solutions of the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, this invention provides a regional load source prediction method based on distributed multi-source data, comprising the following steps:

[0049] Step 1: Collect raw data, obtaining historical parameter data for each distributed power device in the area and historical total load data for the area. The distributed power devices include photovoltaic arrays, wind turbines, energy-saving residential user units, electric vehicle charging piles, and energy storage devices. The historical load data for the area includes total load demand data and total power output data for the area.

[0050] Step 2: Preprocess the historical parameter data of each distributed power device, including data cleaning and normalization.

[0051] Data cleaning includes correcting outliers and filling in missing values ​​in historical parameter data.

[0052] The 3 sigma criterion is used to identify outliers in historical parameter data. The formula is as follows: in, For the first sky The value at time, For the first The average of all recorded times over the day. For the first The standard deviation of all recorded times for each day. For points in the above load formula, these are identified as outliers, and the mean of the records for the corresponding date is used as the outlier correction value. The formula for correcting outliers is as follows: in, For the revised version sky The value at time, For the first The average of all recorded times over the day.

[0053] For missing values ​​in historical parameter data, if the number of missing values ​​is greater than or equal to one-third of the number of recorded time points in a day, the data for that day is discarded; if the number of missing values ​​is less than one-third of the number of recorded time points in a day, the missing values ​​are filled in using data from one time point before and after the missing value on that day, and data from the same time point on the days before and after. The formula for filling in missing values ​​is as follows: in, To fill the gap after the first sky The value at time, and The first sky Time and the sky The value at time, and The first sky Time and the sky The value at time, , The weighting coefficients for the corresponding time and day are respectively. .

[0054] After correcting outliers and filling in missing values ​​in historical parameter data, it is also necessary to normalize the various types of data in each power device to eliminate the influence of different dimensions between different types of data and improve the accuracy of the regional load source prediction model.

[0055] The cleaned historical parameter data is normalized using min-max normalization, as shown in the following formula: in, For the first sky The normalized value at time , After data cleaning sky The value at time, and These are the data cleaning results for the first time. The maximum and minimum values ​​at all recorded times in a given day.

[0056] Step 3: As Figure 2 As shown, a regional load source prediction model is constructed, and trained based on preprocessed historical parameter data and historical total load source data within the region. The regional load source prediction model includes a distributed device terminal and a regional network terminal, and is specifically implemented as follows:

[0057] Step 3.1: Input the preprocessed historical parameter data into the distributed device terminal, perform preliminary fusion through a lightweight coding fusion mechanism, and output the load-source fusion characteristics of the distributed power device, including the following steps:

[0058] Step 3.1.1: Based on the preprocessed historical parameter data of each distributed power device, construct the input sequence of each distributed power device.

[0059] In this embodiment, the data from the preprocessed historical parameter data is taken before the prediction time period. Historical parameter data of each distributed power device at 15-minute intervals (i.e., 12 hours before the prediction period), let the th time point be... Each device has If the data is of type parameter type, then the input sequence is: Each dimension is , length is sequence,

[0060] Step 3.1.2: As Figure 3 As shown, the input sequences are fed into the local coding fusion modules corresponding to the distributed power equipment, and a lightweight coding fusion mechanism is used for preliminary fusion to obtain the load-source fusion features of each distributed power equipment. The specific implementation is as follows:

[0061] (a) Dimension is , length is The input sequences are multiplied by a factor of size 1. The query parameter matrix, key parameter matrix, and value parameter matrix are all of size 1. The query matrix, key matrix, and value matrix;

[0062] (b) The query matrix is ​​weighted and aggregated along the time-series dimension, and the aggregated sequence is normalized to obtain a sequence of size [missing value]. The query sequence;

[0063] (c) Perform element-wise multiplication of the query sequence and the key matrix, and add them along the dimension to obtain a result of size [value missing]. The global query sequence;

[0064] (d) Perform element-wise multiplication of the global query sequence and the value matrix to obtain a result of size . The fusion feature matrix;

[0065] (e) Flatten the fused feature matrix and transform it into a dimension 1 with a length of through a fully connected layer. The load source fusion characteristics are output after the load demand characteristics and power output characteristics are obtained.

[0066] The lightweight coding fusion mechanism is calculated using the following formula: in, These are the query matrix, key matrix, and value matrix of the lightweight fusion mechanism in the local encoding fusion module, respectively. These are the query parameter matrix, key parameter matrix, and value parameter matrix of the lightweight fusion mechanism in the local encoding fusion module, respectively. Given the input sequence, For query sequence, Table normalization function, Indicates aggregation in the direction of time sequence. For global query sequence, This indicates addition along the dimensional direction. As a characteristic of the integration of lotus sources, Indicates a fully connected layer. It means to flatten out.

[0067] Compared to the traditional self-attention mechanism, the lightweight coding fusion mechanism in the local coding fusion module compresses the length of the query matrix through aggregation operations and replaces the matrix multiplication in the original self-attention mechanism with element-wise multiplication, thereby improving the coding fusion efficiency of distributed power equipment and saving computing resources on the distributed equipment side.

[0068] Step 3.2: As Figure 4 and 5 As shown, the load-source fusion characteristics of distributed power equipment and historical total load-source data within the region are input into the regional grid terminal for deep fusion and decoding prediction through an attention mechanism to obtain the total load-source prediction result within the region. The process includes the following steps:

[0069] In this embodiment, as Figure 1 As shown, the core of the regional network terminal is the regional network prediction unit, which consists of a splicing operation, a network encoding and fusion module, and a decoding prediction module.

[0070] The forecast time for load demand and power output is set to 24, with each time point having a time interval of 15 minutes, which means the total regional load value is predicted for the next 6 hours.

[0071] Step 3.2.1: Concatenate the load source fusion features output in Step 3.1 along the dimensional direction to obtain a feature of size [size missing]. The splicing matrix;

[0072] Step 3.2.2: As Figure 4 As shown, the splicing matrix is ​​via A series of cascaded total network coding fusion modules are deeply fused using a self-attention mechanism to obtain a result of size [value missing]. The regional network coding fusion features are specifically implemented as follows;

[0073] (a) The size is The concatenated matrix is ​​used as the input to the first overall network coding fusion module, and multiplied by a value of 1. The query parameter matrix, key parameter matrix, and value parameter matrix are all of size 1. The query matrix, key matrix, and value matrix;

[0074] (b) Multiply the query matrix with the transposed key matrix to obtain the attention matrix, and then normalize it to obtain the normalized attention matrix;

[0075] (c) Perform matrix multiplication on the normalized attention matrix and the value matrix to obtain a matrix of size . Self-attention output;

[0076] (d) The size is The self-attention output is used as the input to the next cascaded total network coding fusion module. Steps (a) to (d) are repeated to perform layer-by-layer fusion until completion. Multiple deep fusions of the coding of a series of cascaded total network coding fusion modules result in a size of... The output is the regional network coding fusion feature;

[0077] The self-attention mechanism in the overall network coding fusion module is calculated using the following formula: in, These are the query matrix, key matrix, and value matrix of the self-attention mechanism in the overall network coding fusion module, respectively. To splice the matrix, These are the query parameter matrix, key parameter matrix, and value parameter matrix of the self-attention mechanism in the overall network coding fusion module, respectively. Table normalization function, This is a feature of regional network coding fusion.

[0078] Since the regional grid has relatively abundant computing resources, the regional grid prediction unit connects multiple grid coding and fusion modules. At the same time, the grid coding and fusion module adopts a computationally intensive self-attention mechanism to perform actual coding and fusion operations, so as to improve the deep fusion effect of historical parameter data of distributed power equipment.

[0079] Step 3.2.3: Using the regional total network coding fusion features and historical total load source data within the region as input to the decoding and prediction module, a supervised attention mechanism is used for decoding and prediction to obtain the total load source prediction result within the region. The specific implementation is as follows:

[0080] (a) Multiply the regional network coding fusion feature by the size. The query parameter matrix yields a matrix of size [missing information]. The query matrix; multiply the historical total load source data within the region by a factor of size . The key parameter matrix and value parameter matrix are obtained as follows: The key matrix and value matrix;

[0081] (b) By using a self-attention mechanism, the query matrix containing the regional total network coding fusion feature information is decoded and interacted with the key matrix and value matrix containing the historical total load source data information within the region, resulting in a matrix of size [missing information]. The decoding matrix;

[0082] (c) Flatten the decoding matrix as A vector of length is transformed through a fully connected layer into... The predicted length, where the first The first forecast value is the predicted total regional load demand value for the next 6 hours, followed by... The predicted value is the predicted total power output of the region within the next 6 hours.

[0083] The supervised attention mechanism in the decoding prediction module is calculated using the following formula: in, These are the query matrix, key matrix, and value matrix of the supervised attention mechanism in the decoding and prediction module, respectively. For the regional network coding fusion characteristics, These are the query parameter matrix, key parameter matrix, and value parameter matrix of the supervised attention mechanism in the decoding and prediction module, respectively. This refers to the historical total load source data for the region. This represents the predicted total load source value within the region. Table normalization function.

[0084] The decoding prediction module employs a unique supervised attention mechanism, introducing historical total load source data within the region as additional auxiliary supervision. This helps alleviate the problem of regional total grid coding fusion feature offset caused by distortion of historical parameter data of some distributed power equipment. While improving the fine fitting capability of the regional load source prediction model, it further ensures the robustness of the model.

[0085] Step 3.3: During the training process of the regional load source prediction model, the model is converged and the model parameters are optimized by using mean square error.

[0086] The loss function for mean squared error is constructed as follows:

[0087] in, The value of the loss function. and These are the loss coefficients for load demand and power output, respectively. In this embodiment, and The values ​​are 0.5 respectively. and The first Forecast values ​​of total regional load demand and total regional power output at specific time points. and The first The actual values ​​of total regional load demand and total regional power output at a given point in time. The value range is 1 to 24. The regional load source prediction model finally outputs 48 predicted values. The first 24 values ​​represent the total regional load demand at the next 24 time points, and the last 24 values ​​represent the total regional power output at the next 24 time points.

[0088] The parameters in the prediction model are iteratively updated using the gradient descent algorithm, thereby continuously reducing the loss function value. When the loss function value hardly changes anymore, it indicates that the regional load source prediction model has converged, the training process is over, and the model parameters are fixed.

[0089] In this embodiment, the prediction model combines the distributed device end and the regional grid end. On the one hand, it makes full use of the massive parameter data of existing and future widely connected new power equipment, which helps to improve the prediction accuracy of total load sources in the region. On the other hand, it realizes the joint calculation of distributed power equipment and regional grid, which solves the problem that the computing resources of regional grid end are difficult to directly process the massive data generated by numerous distributed power equipment.

[0090] Step 4: Input the historical parameter data of each distributed power device in the target area into the regional load source prediction model, and output the predicted value of the total regional load source within the future preset time period to realize the prediction of the total regional load source within the future preset time period.

[0091] In this embodiment, the historical parameter data of each distributed power device in the region for the first 48 time points of the prediction period, i.e. the first 12 hours, are input into the regional load source prediction model. The output load source prediction value with a length of 48 is the predicted value of the total regional load demand and the predicted value of the total regional power output for the next 24 time points (the next 6 hours), thus completing the prediction of the regional load source.

[0092] Another specific embodiment of the present invention provides a regional load source prediction device based on distributed multi-source data, including a processor and a memory. The memory stores a computer program. When the computer program is called and executed by the processor, it implements the regional load source prediction method based on distributed multi-source data as described in any one of claims 1-9.

[0093] All parts not covered in this invention are the same as or implemented using existing technologies.

[0094] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A regional load source prediction method based on distributed multi-source data, characterized in that: Includes the following steps: Step 1: Collect raw data, obtain historical parameter data of each distributed power device in the area and historical total load source data of the area; Step 2: Preprocess the historical parameter data of each distributed power device; Step 3: Construct a regional load source prediction model and train it based on preprocessed historical parameter data and historical total load source data within the region. The regional load source prediction model includes a distributed device terminal and a regional network terminal, specifically implemented as follows: Step 3.1: Input the preprocessed historical parameter data into the distributed device terminal, perform preliminary fusion through a lightweight coding fusion mechanism, and output the load-source fusion characteristics of the distributed power device; Step 3.2: Input the load source fusion characteristics of distributed power equipment and the historical total load source data in the region into the regional grid terminal, and perform deep fusion and decoding prediction through the attention mechanism to obtain the total load source prediction result in the region; Step 4: Input the historical parameter data of each distributed power device in the target area into the regional load source prediction model, and output the predicted value of the total regional load source within the future preset time period to realize the prediction of the total regional load source within the future preset time period.

2. The regional load source prediction method based on distributed multi-source data according to claim 1, characterized in that: The distributed power equipment mentioned in step 1 includes photovoltaic array groups, wind turbine units, energy-saving household user units, electric vehicle charging piles, and energy storage device groups. The historical load source data in the region includes the total load demand data and the total power output data in the region.

3. The regional load prediction method based on distributed multi-source data according to claim 1, characterized in that: The preprocessing of historical parameter data collected from each distributed power device in step 2 includes data cleaning and normalization.

4. The regional load prediction method based on distributed multi-source data according to claim 1, characterized in that: Step 3.1 includes the following steps: Step 3.1.1: Based on the preprocessed historical parameter data of each distributed power device, construct the input sequence for each distributed power device; Step 3.1.2: The input sequences are fed into the local coding fusion modules corresponding to the distributed power equipment, and a lightweight coding fusion mechanism is used for preliminary fusion to obtain the load-source fusion features of each distributed power equipment. The specific implementation is as follows: (a) Dimension is , length is The input sequences are multiplied by a factor of size 1. The query parameter matrix, key parameter matrix, and value parameter matrix are all of size 1. The query matrix, key matrix, and value matrix; (b) The query matrix is ​​weighted and aggregated along the time-series dimension, and the aggregated sequence is normalized to obtain a sequence of size [missing value]. The query sequence; (c) Perform element-wise multiplication of the query sequence and the key matrix, and add them along the dimension to obtain a result of size [value missing]. The global query sequence; (d) Perform element-wise multiplication of the global query sequence and the value matrix to obtain a result of size . The fusion feature matrix; (e) Flatten the fused feature matrix and transform it into a dimension 1 with a length of through a fully connected layer. The output is the result of the fusion of load sources.

5. The regional load prediction method based on distributed multi-source data according to claim 1, characterized in that: The lightweight coding fusion mechanism is calculated using the following formula: in, These are the query matrix, key matrix, and value matrix of the lightweight fusion mechanism in the local encoding fusion module, respectively. These are the query parameter matrix, key parameter matrix, and value parameter matrix of the lightweight fusion mechanism in the local encoding fusion module, respectively. Given the input sequence, For query sequence, Table normalization function, Indicates aggregation in the direction of time sequence. For global query sequence, This indicates addition along the dimensional direction. As a characteristic of the integration of lotus sources, Indicates a fully connected layer. It means to flatten out.

6. The regional load source prediction method based on distributed multi-source data according to claim 1, characterized in that: Step 3.2 includes the following steps: Step 3.2.1: Concatenate the load source fusion features output in Step 3.1 along the dimensional direction to obtain a feature of size [size missing]. The splicing matrix; Step 3.2.2: Matrix splicing via A series of cascaded total network coding fusion modules are deeply fused using a self-attention mechanism to obtain a result of size [value missing]. The regional network coding fusion features are specifically implemented as follows; (a) The size is The concatenated matrix is ​​used as the input to the first overall network coding fusion module, and multiplied by a value of 1. The query parameter matrix, key parameter matrix, and value parameter matrix are all of size 1. The query matrix, key matrix, and value matrix; (b) Multiply the query matrix with the transposed key matrix to obtain the attention matrix, and then normalize it to obtain the normalized attention matrix; (c) Perform matrix multiplication on the normalized attention matrix and the value matrix to obtain a matrix of size . Self-attention output; (d) The size is The self-attention output is used as the input to the next cascaded total network coding fusion module. Steps (a) to (d) are repeated to perform layer-by-layer fusion until completion. The deep fusion of the cascaded total network coding fusion modules yields a result of size [size missing]. The output is the regional network coding fusion feature; Step 3.2.3: Using the regional total network coding fusion features and historical total load source data within the region as input to the decoding and prediction module, a supervised attention mechanism is used for decoding and prediction to obtain the total load source prediction result within the region. The specific implementation is as follows: (a) Multiply the regional network coding fusion feature by the size. The query parameter matrix yields a matrix of size [missing information]. The query matrix; multiply the historical total load source data within the region by a factor of size . The key parameter matrix and value parameter matrix are obtained as follows: The key matrix and value matrix; (b) By using a self-attention mechanism, the query matrix containing the regional total network coding fusion feature information is decoded and interacted with the key matrix and value matrix containing the historical total load source data information within the region, resulting in a matrix of size [missing information]. The decoding matrix; (c) Flatten the decoding matrix as A vector of length is transformed through a fully connected layer into... The predicted length, where the first The first forecast value is the predicted total regional load demand value for a future time period, followed by... Each forecast value represents the predicted total power output of the region over a future period.

7. The regional load prediction method based on distributed multi-source data according to claim 1, characterized in that: Step 3 also includes using mean squared error to converge the model and optimize the model parameters during the training process of the regional load source prediction model.

8. The regional load source prediction method based on distributed multi-source data according to claim 6, characterized in that: Step 3.2.2, the self-attention mechanism in the overall network coding fusion module, the calculation process formula is as follows: in, These are the query matrix, key matrix, and value matrix of the self-attention mechanism in the overall network coding fusion module, respectively. To splice the matrix, These are the query parameter matrix, key parameter matrix, and value parameter matrix of the self-attention mechanism in the overall network coding fusion module, respectively. Table normalization function, This is a feature of regional network coding fusion.

9. A regional load source prediction method based on distributed multi-source data according to claim 6, characterized in that: Step 3.2.3, the supervised attention mechanism in the decoding prediction module, the calculation process formula is as follows: in, These are the query matrix, key matrix, and value matrix of the supervised attention mechanism in the decoding and prediction module, respectively. For the regional network coding fusion characteristics, These are the query parameter matrix, key parameter matrix, and value parameter matrix of the supervised attention mechanism in the decoding and prediction module, respectively. This refers to the historical total load source data for the region. This represents the predicted total load source value within the region. Table normalization function.

10. A regional load prediction device based on distributed multi-source data, comprising a processor and a memory, characterized in that: The memory stores a computer program, which, when called and executed by the processor, implements the regional load source prediction method based on distributed multi-source data as described in any one of claims 1-9.

Images