Power load prediction method and system based on multi-source heterogeneous data, and related equipment

Abstract

The invention is suitable for the technical field of power load prediction, and provides a power load prediction method and system based on multi-source heterogeneous data and related equipment, and the method comprises the steps: collecting the operation data of a power distribution network from a plurality of channels, the operation data of the power distribution network comprising structured data and unstructured data; storing the pre-processed structured data into a preset relational database, and storing the pre-processed unstructured data into a preset non-relational database; performing data fusion on the data in the relational database and the non-relational database through a Bayesian algorithm to obtain target time sequence feature data; inputting the target time sequence feature data as an input sample into the constructed recurrent neural network for model learning training to obtain a trained power load prediction model; and inputting the sequence data obtained in real time into the power load prediction model to generate a target power load prediction value of the power distribution network, thereby effectively improving the accuracy of power load prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a power load forecasting method, system and related equipment based on multi-source heterogeneous data. Background Art

[0002] At present, accurate prediction of power load is of great significance for the planning, operation and management of power systems. On the one hand, from the perspective of power generation, knowing the future power load demand in advance can allow power generation companies to reasonably arrange the start and stop and power generation output of power generation units, and avoid energy waste while meeting the load demand. For example, when it is predicted that the load peak is approaching, the power generation company can start the backup unit in advance to ensure sufficient power supply; and when the load is low, the power generation output can be appropriately reduced to reduce operating costs. In the prior art, the power load forecasting process does not consider the situation of multi-source heterogeneous data, such as only selecting data from certain specific seasons or specific types of users, or only considering a few factors that affect the power load. This leads to the fact that when the forecasting conditions change, the model may not be able to accurately predict, and it is easy to have errors in the power load forecast.

[0003] In view of this, there is a need for a power load forecasting method, system and related equipment based on multi-source heterogeneous data. Summary of the invention

[0004] The embodiments of the present application provide a method, system and related equipment for predicting power load based on multi-source heterogeneous data, which are used to solve the problem of errors in power load prediction.

[0005] A first aspect of an embodiment of the present application provides a method for predicting power load based on multi-source heterogeneous data, comprising:

[0006] Collecting distribution network operation data from multiple channels, wherein the distribution network operation data includes structured data and unstructured data;

[0007] After preprocessing the distribution network operation data, the preprocessed structured data is stored in a preset relational database, and the preprocessed unstructured data is stored in a preset non-relational database;

[0008] The data in the relational database and the data in the non-relational database are fused by using a Bayesian algorithm to obtain target time series feature data;

[0009] The target time series feature data is used as an input sample and input into the constructed recurrent neural network for model learning and training to obtain a trained power load prediction model;

[0010] The sequence data acquired in real time is input into the power load prediction model to generate a target power load prediction value of the distribution network.

[0011] Furthermore, the data in the relational database and the non-relational database are fused by using a Bayesian algorithm to obtain target time series feature data, including:

[0012] Determine a priori probability distribution based on historical power load data;

[0013] Constructing a relational model between power load and data in the relational database and the non-relational database, and determining a likelihood function based on the relational model;

[0014] Calculating a posterior probability distribution based on the prior probability distribution and the likelihood function;

[0015] Target time series feature data is extracted according to the posterior probability distribution.

[0016] Furthermore, the step of constructing a relationship model between the power load and the data in the relational database and the non-relational database includes:

[0017] A regression model is constructed based on the linear relationship between the power load and the data in the relational data and the non-relational database, and the regression coefficient is calculated:

[0018]

[0019] L = α + β D + ∈;

[0020]

[0021] Where: D is any data in a relational database or a non-relational database, L represents the linear relationship between the data vector X and the power load, β is the regression coefficient, α is the intercept, ∈ is the error term, and ∈ obeys the normal distribution N(0,σ 2 ).

[0022] Furthermore, determining the likelihood function based on the relationship model includes:

[0023]

[0024] Where: X1, X2 and X3 are structured data variables in relational databases, x1, x2 and x3 are corresponding specific data values; Y1, Y2 and Y3 are unstructured data variables in non-relational databases, y1, y2 and y3 are corresponding data values. is the normalization constant of the normal distribution, and β1-β6 are the regression coefficients.

[0025] Furthermore, the calculating the posterior probability distribution based on the prior probability distribution and the likelihood function comprises:

[0026]

[0027] Where: P(L) is the prior probability distribution, P(X1=x1,X2=x2,X3=x3,Y1=y1,Y2=y2,Y3=y3) is the normalization constant.

[0028] Furthermore, extracting target time series feature data according to the posterior probability distribution includes:

[0029] Determine a mean and a standard deviation of the posterior distribution according to the posterior probability distribution, and determine a first time series feature vector based on the mean and the standard deviation;

[0030] Extracting a second time series feature vector from the relational database and the non-relational database;

[0031] Target time series feature data is determined based on the first time series feature vector and the second time series feature vector.

[0032] Furthermore, the target time series feature data is used as an input sample and input into the constructed recurrent neural network for model learning and training to obtain a trained power load prediction model, including:

[0033] Dividing the target time series feature data into a training set, a validation set, and a test set, and normalizing the data;

[0034] Build a network architecture based on the normalized data;

[0035] The data in the training set are input into the network in time series order, and the loss function is calculated. Stochastic gradient descent is used for optimization training to obtain a trained power load forecasting model.

[0036] Furthermore, the expression of the loss function is as follows:

[0037]

[0038] Where: y is the actual value of the power load, is the predicted value of power load, and α is the weight coefficient.

[0039] A second aspect of an embodiment of the present application provides a power load forecasting system based on multi-source heterogeneous data, including:

[0040] A distribution network operation data collection unit, used to collect distribution network operation data from multiple channels, wherein the distribution network operation data includes structured data and unstructured data;

[0041] A data storage unit, used for preprocessing the distribution network operation data, storing the preprocessed structured data in a preset relational database, and storing the preprocessed unstructured data in a preset non-relational database;

[0042] A target time series feature data determination unit, configured to perform data fusion on the data in the relational database and the non-relational database by using a Bayesian algorithm to obtain target time series feature data;

[0043] The power load forecasting model determination unit is used to input the target time series feature data as input samples into the constructed recurrent neural network for model learning and training to obtain a trained power load forecasting model;

[0044] The target power load forecast value generating unit is used to input the sequence data acquired in real time into the power load forecast model to generate the target power load forecast value of the distribution network.

[0045] A third aspect of an embodiment of the present application provides a computer device, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein when the processor executes the computer-readable instructions, the steps of the power load forecasting method based on multi-source heterogeneous data as described in any one of the above are implemented.

[0046] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0047] Based on the collected heterogeneous data, the present invention stores structured data and unstructured data in relational databases and non-relational databases respectively, and then uses the Bayesian algorithm to fuse the data in the two types of databases, extract the time series feature data related to the power load, and then input the extracted time series data into the trained prediction model, and finally generate the predicted value of the power load. The present invention uses the Bayesian algorithm to process the data in the two types of databases to obtain rich time series feature data, provide more comprehensive data information for power load prediction, and effectively improve the accuracy of power load prediction by recurrent neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The present invention is a flowchart of an embodiment of a method for power load forecasting based on multi-source heterogeneous data. DETAILED DESCRIPTION

[0049] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example. In addition, the terms "including" and "corresponding to" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0050] Embodiment 1

[0051] The implementation method in this embodiment can be implemented in the system, can be implemented in the server, and can also be implemented in the terminal, and the specific details are not clearly limited. The following will introduce the power load prediction method based on multi-source heterogeneous data in this application from the perspective of system implementation. Figure 1 , the method provided in the embodiment of the present application comprises the following steps:

[0052] S11. Collect distribution network operation data from multiple channels, where the distribution network operation data includes structured data and unstructured data;

[0053] In this embodiment, the distribution network operation data can be collected from the internal systems of the power enterprise, such as the energy management system and the data acquisition and monitoring control system, to obtain the power system operation data, such as power load, power generation data, and power transmission data. It is also possible to collect data through channels such as economic and industrial data, statistical reports issued by government statistical departments, and databases, through censuses and sampling surveys of enterprises and economic activities, and publish relevant economic indicator data after sorting and analysis. It is also possible to obtain some meteorological data and other related data through meteorological departments and data service providers, user feedback channels, and Internet platforms.

[0054] Among the acquired data, structured data generally includes power system operation data, time-related data, and economic and industrial data. The power system operation data includes power load history records, that is, power load data at different time intervals in the past; power generation data, such as the power generation, power generation, and unit operation status of each power plant; transmission and distribution data, such as power flow data of transmission lines, line losses, transformer load rate, oil temperature and other parameters. Time-related data includes time tags such as specific dates, days of the week, and whether it is a holiday; seasonal information, that is, time data divided by season. Economic and industrial data include data such as regional gross domestic product (GDP), industrial added value, industrial structure, and production plan arrangements (such as start time, production plan) and order quantity of specific industrial enterprises.

[0055] Unstructured data generally includes meteorological data and user behavior data. Meteorological data includes descriptions of weather conditions, such as sunny, cloudy, rainy, snowy, etc.; satellite cloud images and weather radar images; and meteorological sensor data, such as wind direction, wind speed, temperature, humidity, air pressure, etc. User behavior data includes user complaints and feedback information, such as user complaints and feedback on power supply quality (such as voltage fluctuations, power outages, etc.), electricity bills, etc.; social media data, such as information posted by users on social media platforms about their experience with electrical equipment, power outages, large-scale events (which may affect power load), etc.

[0056] S12. After preprocessing the distribution network operation data, the preprocessed structured data is stored in a preset relational database, and the preprocessed unstructured data is stored in a preset non-relational database;

[0057] In this embodiment, data preprocessing includes data cleaning, data conversion and data encoding. Data cleaning refers to processing missing values ​​and processing outliers; data conversion is data standardization and normalization; data encoding is classification data encoding and text data encoding.

[0058] Relational databases first design appropriate table structures based on the types and relationships of structured data. For example, for power load history records, create a "power load data table" containing fields such as "timestamp" (time of recording load data), "load value" (actual power load size), "region ID" (indicating the geographical area to which the load data belongs), etc. For power generation data, you can create a "power generation data table" containing fields such as "power plant ID", "generation time", "generation power", and "unit status". Secondly, define the relationship between tables through foreign keys. For example, between the "power load data table" and the "power generation data table", if you want to associate the load data and power generation data at a certain point in time, you can add a "power generation ID" field as a foreign key in the "power load data table" and associate it with the "power generation ID" primary key in the "power generation data table", so that you can establish a logical connection between power load and power generation. Finally, for a small amount of data, you can use the SQL INSERT statement to manually insert data; for a large amount of data, you can use the database's built-in loading tool to store data.

[0059] Non-relational databases include document databases and key-value databases. Specifically, document databases store meteorological data and some user behavior data. First, create databases and collections to store different types of data, and then insert data in JSON format. Key-value databases, taking Redis as an example, Redis stores data in the form of key-value pairs and also supports multiple data structures, such as lists, sets, hashes, etc. If you want to store a series of time series data of user complaints, you can use a list structure.

[0060] S13. Using the Bayesian algorithm to fuse the data in the relational database and the non-relational database, the target time series feature data is obtained;

[0061] Step S13 includes the following:

[0062] 1. Determine the prior probability distribution based on historical power load data;

[0063] First, extract structured data from the relational database, such as power generation, voltage, current, etc. at different time points; extract unstructured data from the non-relational database, such as temperature, humidity, wind speed, etc.; after extracting the data, preprocessing operations such as removing outliers and missing values ​​are required.

[0064] Secondly, we can make a preliminary understanding of the power load based on the historical data we have obtained, and use probability distribution to represent the prior knowledge. For example, we assume that the power load follows a normal distribution without considering new data. Where μ0 and σ0 are the mean and standard deviation estimated based on historical power load data.

[0065] 2. Construct a relational model between power load and data in relational and non-relational databases, and determine the likelihood function based on the relational model;

[0066] Specifically, a regression model is constructed based on the linear relationship between the power load and the data in the relational data and the non-relational database, and the regression coefficient is calculated:

[0067]

[0068] L = α + β D + ∈;

[0069]

[0070] Where: D is any data in a relational database or a non-relational database, L represents the linear relationship between the data vector X and the power load, β is the regression coefficient, α is the intercept, ∈ is the error term, and ∈ obeys the normal distribution N(0,σ 2 ).

[0071] For a given set of data values, such as power generation, voltage, current, temperature, humidity, wind speed; represented as (x1, x2, x3, y1, y2, y3), the likelihood function is:

[0072]

[0073] Where: X1, X2 and X3 are structured data variables in relational databases, x1, x2 and x3 are corresponding specific data values; Y1, Y2 and Y3 are unstructured data variables in non-relational databases, y1, y2 and y3 are corresponding data values. is the normalization constant of the normal distribution, β1-β6 are regression coefficients, The part represents the likelihood of L relative to a given data value, that is, the possibility of L appearing under this condition.

[0074] 3. Calculate the posterior probability distribution based on the prior probability distribution and likelihood function;

[0075] Specifically, according to Bayes' theorem, the posterior probability distribution is:

[0076]

[0077] Where: P(L) is the prior probability distribution, P(X1=x1,X2=x2,X3=x3,Y1=y1,Y2=y2,Y3=y3) is the normalization constant.

[0078] The prior probability distribution and the likelihood function are combined using Bayes' theorem to obtain the posterior probability distribution, which is an updated understanding of the power load after considering the new data. The role of the normalization constant is to ensure that the posterior probability distribution is equal to 1 when integrated over the entire range of values.

[0079] 4. Extract target time series feature data based on the posterior probability distribution.

[0080] Specifically, assume that the posterior probability distribution is still normal According to the results of Bayesian statistics, the mean and standard deviation of the posterior distribution can be calculated by the following formula:

[0081]

[0082] Here, μ1 represents the expected prediction value of power load after integrating new data, Indicates the uncertainty of the prediction.

[0083] Specifically, extracting target time series feature data according to the posterior probability distribution also includes the following steps:

[0084] (1) determining the mean and standard deviation of the posterior distribution according to the posterior probability distribution, and determining the first time series feature vector based on the mean and the standard deviation;

[0085] The mean μ1 of the posterior probability distribution obtained by the above calculation can be used as an important feature, reflecting the expected prediction value of the power load after integrating all the data. As time goes by, a series of means can be obtained to form a time series feature sequence {μ1(t1),μ1(t2),…,μ1(t n )}; standard deviation of the posterior probability distribution It can also be used as a feature to indicate how the uncertainty of the prediction changes over time, also forming a time series That is, the formed time series feature sequence is the first time series feature vector.

[0086] (2) extracting a second time series feature vector from the relational database and the non-relational database;

[0087] The raw data extracted from relational and non-relational databases can also extract time series features. For example, for power generation data, its rate of change at different time points can be calculated. Let x 1,i is time t i The power generation, then the power generation change rate sequence is in (Assume x 1,0 As an example, for meteorological data, such as temperature, the daily variation of temperature can be calculated.1,i is time t i For the temperature data within a day, the daily variation series can be calculated in t i,max and t i,min are the highest and lowest temperatures of the day, respectively. That is, the power generation change rate sequence and the daily change amplitude sequence are the second time series feature vectors.

[0088] (3) Determine target time series feature data based on the first time series feature vector and the second time series feature vector.

[0089] The time series features extracted from the posterior probability distribution (first time series feature vector) and the original data (second time series feature vector) are combined to form a comprehensive feature vector. For example, the target time series feature data can be expressed as

[0090] S14. The target time series feature data is used as an input sample and input into the constructed recurrent neural network for model learning and training to obtain a trained power load forecasting model;

[0091] S15. Input the sequence data acquired in real time into the power load forecasting model to generate a target power load forecast value for the distribution network.

[0092] Step S14 includes the following:

[0093] 1. Divide the target time series feature data into training set, validation set and test set, and normalize the data;

[0094] Specifically, the extracted time series feature data is divided into training set, validation set and test set. Generally speaking, the training set is used to train the network, the validation set is used to adjust the network's hyperparameters, and the test set is used to evaluate the network's final performance. The division ratio is 70% for training, 15% for validation, and 15% for testing. Normalization is to map the data to a smaller interval to improve the efficiency and stability of network training.

[0095] 2. Build a network architecture based on the normalized data;

[0096] First, determine the number of nodes in the input layer. The dimension of the input time series feature data determines the number of nodes in the input layer. For example, if the time series feature vector contains m features such as the mean, standard deviation, rate of change of power generation, daily temperature change amplitude, etc. of the posterior probability distribution of power load, then the number of input layer nodes is m. Secondly, select the appropriate number of hidden layers and the number of neurons in each hidden layer. You can start with a simple architecture, such as one hidden layer, and then gradually increase the number of hidden layers or the number of neurons to observe the performance of the model on the validation set. Assume that one hidden layer is set and the number of neurons is h. Finally, the output layer sets a node to output the predicted power load.

[0097] 3. Input the data in the training set into the network in time series order, calculate the loss function, and use stochastic gradient descent for optimization training to obtain a trained power load forecasting model.

[0098] The training data is input into the network in time series order, and at each time step t, the hidden layer state is calculated. where x t is the input feature vector at time t, that is, the normalized time series feature data, U and W are weight matrices, b is the bias term, f is the activation function, and h t―1 is the hidden layer state of the previous time step. Calculate the output layer output Where V is the output layer weight matrix, c is the output layer bias term, and g is the output layer activation function. The loss function is calculated based on the predicted value and the actual value, and the loss function is defined as follows:

[0099]

[0100] Where: y is the actual value of the power load, is the predicted value of power load, and α is the weight coefficient.

[0101] Finally, the stochastic gradient descent optimization algorithm is selected to update the network weights and bias terms according to the gradient back propagation of the loss function. This process is repeated until the predetermined number of training rounds (such as 100 rounds) is reached or the loss function converges (such as the value of the loss function changes less than a set threshold in several consecutive rounds of training).

[0102] When the model performs well on the test set, it can be applied to actual power load forecasting. By inputting the latest normalized time series feature data into the trained model, the predicted value of future power load can be obtained. At the same time, as new data is continuously generated, the model can be regularly updated and retrained to improve the accuracy of the forecast.

[0103] Embodiment 2

[0104] An embodiment of a power load forecasting system based on multi-source heterogeneous data in the present invention comprises the following steps:

[0105] A distribution network operation data collection unit, used to collect distribution network operation data from multiple channels, the distribution network operation data including structured data and unstructured data;

[0106] A data storage unit is used to perform data preprocessing on the distribution network operation data, store the preprocessed structured data in a preset relational database, and store the preprocessed unstructured data in a preset non-relational database;

[0107] A target time series feature data determination unit is used to perform data fusion on the data in the relational database and the non-relational database through a Bayesian algorithm to obtain target time series feature data;

[0108] The power load forecasting model determination unit is used to input the target time series feature data as input samples into the constructed recurrent neural network for model learning and training to obtain a trained power load forecasting model;

[0109] The target power load forecast value generating unit is used to input the sequence data acquired in real time into the power load forecast model to generate the target power load forecast value of the distribution network.

[0110] For the specific definition of the system, please refer to the definition of the method above, which will not be repeated here. Each module in the above system can be implemented in whole or in part by software, hardware and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0111] Embodiment 3

[0112] The present invention provides a computer device, comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, the steps of the above method are implemented.

[0113] Those of ordinary skill in the art will appreciate that the units of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0114] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored. In addition, each functional unit in each embodiment of the present invention can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units.

[0115] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nlyMemory), random access memory (RAM, RandomAccessMemory), mobile hard disk, magnetic disk or optical disk, etc., which can store program code.

[0116] It can be understood that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A method for predicting power load based on multi-source heterogeneous data, characterized in that: include: Collecting distribution network operation data from multiple channels, wherein the distribution network operation data includes structured data and unstructured data; After preprocessing the distribution network operation data, the preprocessed structured data is stored in a preset relational database, and the preprocessed unstructured data is stored in a preset non-relational database; The data in the relational database and the data in the non-relational database are fused by using a Bayesian algorithm to obtain target time series feature data; The target time series feature data is used as an input sample and input into the constructed recurrent neural network for model learning and training to obtain a trained power load prediction model; The sequence data acquired in real time is input into the power load prediction model to generate a target power load prediction value of the distribution network.

2. The method for predicting power load based on multi-source heterogeneous data according to claim 1, characterized in that: The method of fusing the data in the relational database and the data in the non-relational database by using a Bayesian algorithm to obtain target time series feature data includes: Determine a priori probability distribution based on historical power load data; Constructing a relational model between power load and data in the relational database and the non-relational database, and determining a likelihood function based on the relational model; Calculating a posterior probability distribution based on the prior probability distribution and the likelihood function; Target time series feature data is extracted according to the posterior probability distribution.

3. The method for predicting power load based on multi-source heterogeneous data according to claim 2 is characterized in that: The step of constructing a relationship model between the power load and the data in the relational database and the non-relational database includes: A regression model is constructed based on the linear relationship between the power load and the data in the relational data and the non-relational database, and the regression coefficient is calculated: L = α + β D + ∈; Where: D is any data in a relational database or a non-relational database, L represents the linear relationship between the data vector X and the power load, β is the regression coefficient, α is the intercept, ∈ is the error term, and ∈ obeys the normal distribution N(0,σ 2 ).

4. The method for predicting power load based on multi-source heterogeneous data according to claim 3 is characterized in that: Determining the likelihood function based on the relationship model includes: Where: X1, X2 and X3 are structured data variables in relational databases, x1, x2 and x3 are corresponding specific data values; Y1, Y2 and Y3 are unstructured data variables in non-relational databases, y1, y2 and y3 are corresponding data values. is the normalization constant of the normal distribution, and β1-β6 are the regression coefficients.

5. The method for predicting power load based on multi-source heterogeneous data according to any one of claims 1 to 4, characterized in that: The calculating the posterior probability distribution based on the prior probability distribution and the likelihood function comprises: Where: P(L) is the prior probability distribution, P(x1=x1,X2=x2,X3=x3,Y1=y1,Y2=y2,Y3=y3) is the normalization constant.

6. The method for power load forecasting based on multi-source heterogeneous data according to claim 5 is characterized in that: The extracting target time series feature data according to the posterior probability distribution includes: Determine a mean and a standard deviation of the posterior distribution according to the posterior probability distribution, and determine a first time series feature vector based on the mean and the standard deviation; Extracting a second time series feature vector from the relational database and the non-relational database; Target time series feature data is determined based on the first time series feature vector and the second time series feature vector.

7. The method for power load forecasting based on multi-source heterogeneous data according to claim 1, characterized in that: The target time series feature data is used as an input sample and input into the constructed recurrent neural network for model learning and training to obtain a trained power load prediction model, including: Dividing the target time series feature data into a training set, a validation set, and a test set, and normalizing the data; Build a network architecture based on the normalized data; The data in the training set are input into the network in time series order, and the loss function is calculated. Stochastic gradient descent is used for optimization training to obtain a trained power load forecasting model.

8. The method for power load forecasting based on multi-source heterogeneous data according to claim 7 is characterized in that: The expression of the loss function is as follows: Where: y is the actual value of the power load, is the predicted value of power load, and α is the weight coefficient.

9. A power load forecasting system based on multi-source heterogeneous data, characterized in that: include: A distribution network operation data collection unit, used to collect distribution network operation data from multiple channels, wherein the distribution network operation data includes structured data and unstructured data; A data storage unit, used for preprocessing the distribution network operation data, storing the preprocessed structured data in a preset relational database, and storing the preprocessed unstructured data in a preset non-relational database; A target time series feature data determination unit, configured to perform data fusion on the data in the relational database and the non-relational database by using a Bayesian algorithm to obtain target time series feature data; The power load forecasting model determination unit is used to input the target time series feature data as input samples into the constructed recurrent neural network for model learning and training to obtain a trained power load forecasting model; The target power load forecast value generating unit is used to input the sequence data acquired in real time into the power load forecast model to generate the target power load forecast value of the distribution network.

10. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that: When the processor executes the computer-readable instructions, the steps of the power load forecasting method based on multi-source heterogeneous data as described in any one of claims 1 to 8 are implemented.

Images