An intelligent power load prediction method fusing service logic and deep learning

Abstract

The application discloses a power consumption load intelligent prediction method fusing business logic and deep learning, extracts an optimal typical day as business logic prior through similar day retrieval, constructs a mixed feature input, and combines a BiLSTM timing network, so that the load fluctuation law in nonlinear mutation scenes such as holiday adjustment, temporary maintenance of large users and photovoltaic reverse feedback can be effectively captured. The analogy calculation logic of an expert business person is converted into prior knowledge constraints of a neural network, so that the model output not only depends on data driving, but also can give reasonable basis conforming to business common sense, thereby effectively making up for the defect that a pure black box model is difficult to understand, improving the trust degree of power dispatch personnel on the prediction result, and providing more reliable decision support for power grid safe operation, power generation plan arrangement and power market transaction.

Description

Technical Field

[0001] This application relates to the fields of power system automation and artificial intelligence application technology, specifically to a method for intelligent prediction of electricity load that integrates business logic and deep learning. Background Technology

[0002] Traditional electricity consumption forecasting methods mainly include statistical methods such as time series analysis and regression analysis, as well as deep learning methods based on artificial neural networks. While purely data-driven model methods can capture the temporal patterns of historical data, they lack sufficient utilization of business logic and struggle to understand and respond to sudden changes in electricity consumption caused by business rules. For example, special scenarios such as temporary maintenance by large users, holiday work schedule adjustments, and reverse power feedback from photovoltaic users often cause pure data model predictions to fail. Furthermore, conventional models are usually designed for general users and lack refined modeling for special entities such as large industrial users and distributed photovoltaic users, resulting in overall prediction accuracy being greatly affected by local anomalies.

[0003] Therefore, there is an urgent need for an intelligent electricity consumption forecasting method that can combine expert business logic with the advantages of deep learning, possessing both data mining capabilities and adaptability to business rules and interpretability. Summary of the Invention

[0004] This application aims to improve the applicability of electricity consumption forecasting models to specific entities and their ability to understand electricity consumption anomalies caused by business rules, thereby improving the accuracy of electricity consumption forecasting. It provides an intelligent electricity load forecasting method that integrates business logic and deep learning.

[0005] To achieve this objective, the following technical solution is adopted in this application:

[0006] A method for intelligent electricity load forecasting that integrates business logic and deep learning is provided, including the following steps:

[0007] S1, construct a multi-dimensional feature space for electricity consumption data to transform the raw electricity consumption data into feature vectors that include business semantics;

[0008] S2, based on the business logic of similar day retrieval, extracts the optimal typical day and determines the typical benchmark electricity load curve for the target prediction day;

[0009] S3, based on the multidimensional feature space, constructs a hybrid input neural network model as a prediction engine, and outputs the predicted electricity load curve for the target day.

[0010] Preferably, the method for constructing the multidimensional feature space of electricity consumption data in step S1 specifically includes the following steps:

[0011] S11, After preprocessing the historical electricity consumption data collected at a preset frequency, mark the date attribute characteristics and / or meteorological characteristics;

[0012] S12, Construct a user knowledge base, including a user list and / or user business information; the users in the user list include photovoltaic power generation users, and the user business information includes one or more of the following: photovoltaic installed capacity, maintenance plan, and production schedule.

[0013] Preferably, the method for preprocessing historical electricity consumption data with abnormal mutations and missing data is expressed as follows:

[0014]

[0015] in, Represents the window length; Indicates the length of the window radius; Indicates the first in the window One data collection point; Indicates the first The original electricity consumption values ​​of each collection point; This is the pre-processed power consumption value.

[0016] Preferably, the method for extracting the optimal typical day based on the business logic of similar day retrieval includes the following steps:

[0017] S21, based on date hard constraint matching, select the historical day that is completely consistent with and closest to the date attribute of the prediction target day as the candidate reference day;

[0018] S22, based on meteorological soft constraint matching, selects the historical day with the smallest deviation from the meteorological characteristics of the predicted target day among the candidate reference days as the optimal typical day.

[0019] Preferably, the method for calculating the deviation of meteorological characteristics between the candidate reference date and the predicted target date includes the following steps:

[0020] S221, predicting the target date and the Meteorological feature matrix is ​​constructed from meteorological data collected at several electricity consumption data collection times for each candidate reference day:

[0021] S222, dimensionless processing is performed on each eigenvalue in the meteorological feature matrix;

[0022] S223, for each time step The predicted target date will be in this Meteorological characteristic sequence at time As a reference sequence, the first Meteorological characteristic sequence of candidate reference days As a comparison sequence, calculate each Correlation coefficient at time ; This represents the dimensionless value of the z-th meteorological feature at time t on the predicted target day; This represents the dimensionless value of the z-th meteorological feature at time t for the i-th candidate reference day;

[0023] S224, calculate the... Each of the candidate reference days Correlation coefficient at time The mean of, as the first The comprehensive correlation between the meteorological characteristic sequences of the candidate reference days and the meteorological characteristic sequences of the predicted target days. and to Sort in descending order and filter out the items before sorting. The candidate reference date is selected as the optimal typical date.

[0024] Preferably, the correlation coefficient The calculation method is expressed as follows:

[0025]

[0026]

[0027]

[0028]

[0029] in, Representing the The time of each candidate reference date The absolute difference of each meteorological feature;

[0030] and These are the two minimum and two maximum differences across all times, all meteorological characteristics, and all candidate reference days;

[0031] To take the minimum value among all candidate reference days; To take the minimum value among all meteorological characteristics; To find the minimum value over all times.

[0032] Represents the resolution coefficient.

[0033] Preferably, in step S2, the method for determining the typical baseline electricity load curve for the predicted target day based on the optimal typical day is as follows:

[0034] set up For the k-th optimal typical day, its 48-point load data is as follows: .

[0035] like Predicting the daily electricity load baseline curve The electricity load data for the optimal typical day is directly taken;

[0036] like Predicting the daily electricity load baseline curve The electricity load data for the two optimal typical days are generated by using either equal-weighted average or weighted average.

[0037] The typical baseline electricity load curve uses the time of electricity consumption data collection as the horizontal axis and the electricity consumption data as the vertical axis.

[0038] Preferably, in step S3, the method for constructing the hybrid input of the model based on the multidimensional feature space is as follows:

[0039] Historical time-series features, static covariant features, and business logic features carried in the multidimensional feature space are extracted, and then standardized and concatenated along the feature dimension. The concatenated result is used as the mixed input of the model.

[0040] Historical time-series characteristics include historical load data for L consecutive days prior to the forecast target date;

[0041] Static covariant features include date features and / or weather forecast features. Date features include one or more of the following: date attribute identifier, month, weekday, and time of the predicted target date. Weather forecast features include one or more of the following: temperature, humidity, cloud cover, and irradiance of the predicted target date.

[0042] Business logic features include predicting the electricity load forecast curve for the target day;

[0043] Feature standardization processing includes differential feature processing of extracted categorical and numerical features, including:

[0044] For unordered category features, one-hot encoding is used to convert them into binary numerical vectors;

[0045] For ordered category features, trigonometric function encoding is used to expand them into continuous periodic values;

[0046] For continuous numerical features, Z-value standardization is used to transform continuous numerical values ​​into dimensionless standard values.

[0047] Preferably, in step S3, the method for constructing the neural network model includes the following steps:

[0048] The constructed hybrid input and the measured electricity load data collected at various times on the current target day are used as training samples. The neural network model is trained through a BiLSTM bidirectional long short-term memory network. The training process is expressed as follows:

[0049]

[0050]

[0051]

[0052] in, and Represent The output of the LSTM layer is fed back from the previous and next time steps; represent Mixed feature input at time step; and These represent the weight matrices from the forward and backward LSTM layers to the output layer, respectively. Represents the bias term of the output layer; express The model output at time step; and Represent The output of the LSTM layer is fed back from the previous and next time steps;

[0053] The above training process is supervised by the Huber loss function, and the model's gradient updates and weight configurations are based on the loss function values. When continuous... When the rate of decrease of the loss function value is less than the preset rate of decrease threshold, stop model training, save the model parameters, and obtain the target neural network model;

[0054] The gradient update and weight configuration process calculates the model parameters for each training round. gradient of the loss function The model parameter weights are updated based on the gradient direction, and the calculation formula is as follows:

[0055]

[0056]

[0057] in, These represent the parameters that the model needs to be trained, including the weight matrix and bias terms; Represents the learning rate;

[0058] The value of the loss function;

[0059] Indicates the first The parameters required for training the round model;

[0060] Indicates the first The parameters required for training the round model.

[0061] This application has the following beneficial effects:

[0062] 1. By integrating business logic (such as similar day retrieval and special user knowledge base) with deep learning models, it can effectively address the prediction failure problem caused by nonlinear sudden change scenarios such as holiday work schedule adjustments, temporary maintenance of large users, and photovoltaic reverse feed. It significantly reduces the prediction error of pure data-driven models in these scenarios, and keeps the overall prediction accuracy stable in environments with a high proportion of renewable energy access.

[0063] 2. By using expert business logic (such as typical day matching and post-processing of business rules) as prior knowledge to constrain the neural network, the model's prediction output not only depends on data-driven approaches but also provides reasonable evidence that conforms to common business sense (such as adjustments based on similar daily load curves and user production plans). This enhances the power dispatchers' confidence in the prediction results and facilitates actual business deployment and application. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments of this application will be briefly described below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart illustrating the implementation steps of an intelligent power load prediction method that integrates business logic and deep learning, as provided in an embodiment of this application.

[0066] Figure 2 This is a diagram of the neural network structure used for model training in this application. Detailed Implementation

[0067] The technical solution of this application will be further described below with reference to the accompanying drawings and specific embodiments.

[0068] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of this application, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0069] In the accompanying drawings of the embodiments of this application, the same or similar reference numerals correspond to the same or similar components. In the description of this application, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting this application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0070] In the description of this application, unless otherwise expressly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0071] This application provides a method for intelligent electricity load prediction that integrates business logic and deep learning, such as... Figure 1 As shown, the steps include:

[0072] S2, construct a multi-dimensional feature space for electricity consumption data to transform the raw electricity consumption data into feature vectors that include business semantics;

[0073] Specifically, the method for constructing the multidimensional feature space of electricity consumption data includes the following steps:

[0074] S11. Extract historical electricity consumption data of the target area for at least one year. The sampling frequency is preferably 48 points per day (data is collected once every half hour). For abnormal changes and missing data, the forward and backward filling method is used to smooth the completion.

[0075] Specifically, based on the 3σ criterion, if the first... Electricity consumption at each data collection point satisfy If a data point is missing, it is marked as an anomalous mutation point. For each anomalous mutation point and missing point, the mean of the values ​​within the preceding and following windows is used to replace the original value, resulting in a smoothed value. for:

[0076]

[0077] in, This represents the average of the total electricity consumption data for that day. The standard deviation of the total electricity consumption data for that day; Represents the window length; Indicates the length of the window radius; Indicates the first in the window One data collection point; Indicates the first Electricity consumption values ​​at each data collection point;

[0078] Then, the date attribute features are labeled on the smoothed and completed historical electricity consumption data. The labeling method is as follows:

[0079] Based on the national statutory holiday schedule, each day is assigned a date type label, including ordinary workday, ordinary weekend, statutory holiday, and workday adjusted for work.

[0080] And / or, annotate the smoothed and completed historical electricity consumption data with meteorological characteristics, using the following annotation method:

[0081] Acquire historical meteorological data for the target area, including one or more of temperature, humidity, cloud cover, wind speed, and irradiance. The preferred sampling frequency for meteorological data is 48 points per day, i.e., once every half hour. Then, each collected historical meteorological data point is assigned as a meteorological feature label to the corresponding historical electricity consumption data.

[0082] For example, if electricity consumption data is collected at 11:00 AM on January 1, 2025, and the meteorological data at that time is such that the temperature is 20℃ and the day is Monday, then the date attribute of the electricity consumption data at that time is Monday, and the meteorological characteristic is a temperature of 20℃.

[0083] S12, Construct a user knowledge base, including a user list and / or user business information. Users in this application include photovoltaic power generation users and large electricity consumers (e.g., defining a large electricity consumer as one with an annual electricity consumption exceeding 5 million kWh). User business information includes one or more of the following: photovoltaic installed capacity, maintenance plans, and production scheduling plans.

[0084] In this application, the date attribute features, meteorological features, and user knowledge base labeled on historical electricity consumption data constitute a multidimensional feature space for historical electricity consumption data.

[0085] After completing the construction of the multidimensional feature space for electricity consumption data through the above step S1, as follows: Figure 1 As shown, the steps of the intelligent power load prediction method that integrates business logic and deep learning provided in this embodiment are as follows:

[0086] S2, based on the business logic of similar day retrieval, extracts the optimal typical day and determines the typical benchmark electricity load curve for the target prediction day, specifically including the following steps:

[0087] S21, based on date hard constraint matching, select the historical day that is completely consistent with and closest to the date attribute of the prediction target day as the candidate reference day;

[0088] For example, if the predicted target day is a regular workday, historical days with the same regular workday and weekday attributes as the predicted target day are preferred as candidate reference days; if the predicted target day is a regular weekend, historical days with the same regular weekend and weekday attributes as the predicted target day are preferred as candidate reference days; if the predicted target day is a statutory holiday, the same statutory holidays in the past 1-2 years are preferred as candidate reference days; if the predicted target day is a workday adjusted for work, the same workday adjusted for work in the past 1-2 years are preferred as candidate reference days.

[0089] S22, based on meteorological soft constraint matching, selects the historical day with the smallest deviation from the meteorological characteristics of the predicted target day among the candidate reference days as the optimal typical day.

[0090] In this embodiment, the method for calculating the deviation of meteorological characteristics between the candidate reference date and the predicted target date includes the following steps:

[0091] S221, predicting the target date and the Meteorological feature matrix is ​​formed by meteorological data from 48 electricity consumption data collection times on each candidate reference day;

[0092]

[0093]

[0094] , Representing the predicted target date and the [number]th [day], respectively. Meteorological characteristic matrix of candidate reference days;

[0095] Each row Indicates time of Each meteorological characteristic value; each row Indicates the first The time of each candidate reference date of One meteorological characteristic value;

[0096] S222, Dimensionless processing is performed on each eigenvalue in the meteorological feature matrix, using the following method:

[0097]

[0098] This represents the meteorological characteristic value after dimensionless processing; Represents the original meteorological characteristic value;

[0099] S223, for each time step The predicted target date will be in this Meteorological characteristic sequence at time As a reference sequence, the first Meteorological characteristic sequence of candidate reference days As a comparison sequence, calculate each Correlation coefficient at time ; This represents the dimensionless value of the z-th meteorological feature at time t on the predicted target day; This represents the dimensionless value of the z-th meteorological feature at time t for the i-th candidate reference day;

[0100] In this embodiment, the correlation coefficient The calculation method is expressed as follows:

[0101]

[0102]

[0103]

[0104]

[0105] in, Representing the The time of each candidate reference date The absolute difference of each meteorological feature;

[0106] and These are the two minimum and two maximum differences across all times, all meteorological characteristics, and all candidate reference days;

[0107] To take the minimum value among all candidate reference days; To take the minimum value among all meteorological characteristics; To find the minimum value over all times.

[0108] This represents the resolution coefficient, which is usually taken as 0.5.

[0109] S224, calculate the... Each of the candidate reference days Correlation coefficient at time The mean of, as the first The comprehensive correlation between the meteorological characteristic sequences of the candidate reference days and the meteorological characteristic sequences of the predicted target days. and to Sort in descending order and filter out the items before sorting. The candidate reference date is selected as the optimal typical date.

[0110]

[0111] In this embodiment, temperature, humidity, and irradiance are taken as meteorological characteristics.

[0112] The following example illustrates the calculation process for the optimal typical day:

[0113] For example, the meteorological feature matrix for predicting the target day is:

[0114]

[0115] The meteorological characteristic matrix of the candidate reference day is as follows:

[0116]

[0117] The dimensionless meteorological feature matrix is ​​obtained by range normalization:

[0118]

[0119]

[0120] yes The dimensionless expression of .

[0121] Then, calculate and The absolute difference between each meteorological feature in these two meteorological feature matrices at each time point :

[0122]

[0123] Iterate through all elements in the absolute difference matrix to determine the minimum and maximum differences at both levels:

[0124]

[0125]

[0126] Calculate the correlation coefficient at each time point. :

[0127]

[0128] Similarly, , , ,

[0129] The overall correlation degree is obtained by averaging the correlation coefficients at all times. :

[0130]

[0131] In step S2, the method for determining the typical baseline electricity load curve for the target prediction day based on the optimal typical day is as follows:

[0132] set up For the k-th optimal typical day, its 48-point load data is as follows: .

[0133] like Predicting the daily electricity load baseline curve The electricity load data for the optimal typical day is directly taken;

[0134] like Predicting the daily electricity load baseline curve The electricity load data for the two optimal typical days are generated by using either equal-weighted average or weighted average.

[0135] The typical baseline electricity load curve uses the time of electricity consumption data collection as the horizontal axis and the electricity consumption data as the vertical axis.

[0136] After determining the typical baseline electricity load curve for the target forecast date in step S2, as follows: Figure 1 As shown, the steps of the intelligent power load prediction method integrating business logic and deep learning provided in this application are as follows:

[0137] S3, based on the multidimensional feature space, constructs a hybrid input neural network model as a prediction engine, and outputs the predicted electricity load curve for the target day;

[0138] In this embodiment, the method for constructing the hybrid input of the model based on the multidimensional feature space is as follows:

[0139] Historical time-series features, static covariant features, and business logic features carried in the multidimensional feature space are extracted, and then standardized and concatenated along the feature dimension. The concatenated result is used as the mixed input of the model.

[0140] The historical time series features are preferably historical load data (electricity consumption data) for L consecutive days before the target date of prediction, and the window length L is preferably 7 or 14 to ensure that the model input can fully cover the intra-week fluctuation pattern of the cycle;

[0141] Static covariance features include date features and / or weather forecast features. The preferred date features are the date attribute identifier, month, weekday, and time of the target date. The preferred weather forecast features are the temperature, humidity, cloud cover, and irradiance of the target date. The preferred business logic features are the electricity load forecast curve for the target date.

[0142] Feature standardization processing includes differential feature processing of extracted categorical and numerical features, including:

[0143] For unordered categorical features such as weekdays, regular rest days, and statutory holidays, one-hot encoding is used to convert them into binary numerical vectors. For example, weekdays are mapped as... Ordinary rest days are mapped as Statutory holidays are mapped as .

[0144] For ordered categorical features such as day of the week, time of day, and month, sinusoidal encoding is used to expand them into continuous periodic values. The calculation formula is as follows:

[0145]

[0146]

[0147] In the formula, The numerical value represents the feature corresponding to the ordered category. This represents the total number of categories for an ordered categorical feature. Taking the weekday feature as an example... They correspond to Monday through Sunday in sequence. It is 7.

[0148] For continuous numerical characteristics such as historical electricity consumption, electricity load baseline curves, and meteorological data, Z-score standardization is used to transform continuous values ​​into dimensionless standard values ​​with a mean of 0 and a standard deviation of 1. The calculation formula is as follows:

[0149]

[0150] In the formula, The first dimensionless quantity A continuous numerical eigenvalue, Representing all The average value, Represents standard deviation.

[0151] In step S3, the neural network model constructed based on the hybrid input is not limited to a specific neural network architecture, but preferably adopts a neural network model that can handle long temporal dependencies and has interpretability. In this embodiment, a bidirectional long short-term memory network (BiLSTM) is selected as the core testing engine.

[0152] Bidirectional Long Short-Term Memory Network is an optimized model evolved from the Long Short-Term Memory Network (LSTM) architecture. Its core breakthrough lies in the adoption of a symmetrical bidirectional structure in the hidden layers. By integrating forward LSTM layers and backward LSTM layers, the model can simultaneously capture past historical information and future trend information in time series. Figure 2 The architecture diagram of the BiLSTM network used in this application is shown. The calculation formula for the BiLSTM model is:

[0153]

[0154]

[0155]

[0156] in, and Represent The output of the LSTM layer is fed back from the previous and next time steps; represent Mixed feature input at time step; and These represent the weight matrices from the forward and backward LSTM layers to the output layer, respectively. Represents the bias term of the output layer; express The model output at time step; and Represent Forward and backward outputs of the LSTM layer at each time step

[0157] The hybrid feature dataset used to train the deep neural network model consists of training sample pairs composed of standardized multidimensional hybrid feature inputs and corresponding measured electricity load outputs. Specifically, a single training sample pair uses the hybrid feature vector, which integrates historical load sequences, static covariance features, and business logic features in step S3, as the model input, and uses the measured electricity load data corresponding to the 48 sampling times of the entire prediction day as the output labels for supervised learning.

[0158] The training process is supervised by the Huber loss function. The model's gradient is updated and weights are configured based on the loss function value. If the performance does not improve after 30 consecutive training iterations, the training is stopped, and the parameters of the best-performing model are saved to obtain the target neural network model.

[0159] The Huber loss function is a robust loss function that combines the advantages of mean squared error (MSE) and mean absolute error (MAE), which can effectively improve the model's sensitivity to extreme application values. The calculation formula is as follows:

[0160]

[0161] in, This represents the actual amount of electricity consumed; This represents the predicted electricity consumption value; This represents the positive threshold parameter, which is usually set to 1.

[0162] The gradient update and weight configuration process involves calculating the parameters of the model in each training round. gradient of the loss function The model parameter weights are updated based on the gradient direction, and the calculation formula is as follows:

[0163]

[0164]

[0165] in, These represent the parameters that the model needs to be trained, including the weight matrix and bias terms; This represents the learning rate.

[0166] Finally, the trained neural network model is used to output the predicted electricity load curve for the target day.

[0167] Specifically, the trained neural network model in this application is persistently stored in the form of a model architecture file and a model parameter weight file. In actual deployment applications, the model architecture is first loaded into an embedded computer to construct a network computation graph. Then, the model parameter file is read and the weight values ​​are loaded into the constructed model structure, completing the model loading. Next, the hybrid features constructed for predicting the target day are input into the neural network model, and the model outputs the predicted electricity load data for the target day, preferably plotted as an electricity load prediction curve.

[0168] For business scenarios that are not fully covered by neural network models, business rules are introduced to perform post-processing corrections on special user loads, achieving a deep integration of expert knowledge and data models.

[0169] Specifically, for users with photovoltaic power generation, it is preferable to estimate the photovoltaic output curve based on the irradiance data and photovoltaic installed capacity from the weather forecast, and then correct the electricity load forecast curve for the target date. For large electricity users, the base load portion of the electricity load forecast curve for the target date can be linearly scaled and corrected based on the adjustment coefficient of their declared production plan.

[0170] In summary, this application extracts the optimal typical day as a priori for business logic through similar day retrieval, constructs a hybrid feature input, and combines it with time series networks such as BiLSTM. This effectively captures load fluctuation patterns under nonlinear abrupt changes such as holiday work schedule adjustments, temporary maintenance of large users, and photovoltaic reverse feedback. By transforming the analogical calculation logic of expert business personnel into prior knowledge constraints of neural networks, the model output not only relies on data-driven approaches but also provides reasonable justifications consistent with common business sense. This effectively compensates for the incomprehensibility of pure black-box models, enhances the trust of power dispatchers in the prediction results, and provides more reliable decision support for grid safety operation, power generation planning, and electricity market transactions.

[0171] It should be stated that the above-described specific embodiments are merely preferred embodiments and technical principles applied in this application. Those skilled in the art should understand that various modifications, equivalent substitutions, and variations can be made to this application. However, such variations, as long as they do not depart from the spirit of this application, should be within the scope of protection of this application. Furthermore, some terminology used in this application's specification and claims is not limiting but merely for ease of description.

Claims

1. A method for intelligent forecasting of electricity load that integrates business logic and deep learning, characterized in that, Including the following steps: S1, construct a multi-dimensional feature space for electricity consumption data to transform the raw electricity consumption data into feature vectors that include business semantics; S2, based on the business logic of similar day retrieval, extracts the optimal typical day and determines the typical benchmark electricity load curve for the target prediction day; S3, based on the multidimensional feature space, constructs a hybrid input neural network model as a prediction engine, and outputs the predicted electricity load curve for the target day.

2. The intelligent power load prediction method integrating business logic and deep learning according to claim 1, characterized in that, The method for constructing the multidimensional feature space of electricity consumption data in step S1 specifically includes the following steps: S11, After preprocessing the historical electricity consumption data collected at a preset frequency, mark the date attribute characteristics and / or meteorological characteristics; S12, Construct a user knowledge base, including a user list and / or user business information; the users in the user list include photovoltaic power generation users, and the user business information includes one or more of the following: photovoltaic installed capacity, maintenance plan, and production schedule.

3. The intelligent load forecasting method integrating business logic and deep learning according to claim 2, characterized in that, The method for preprocessing historical electricity consumption data that contains abnormal mutations and missing data is expressed as follows: in, Represents the window length; Indicates the length of the window radius; Indicates the first in the window One data collection point; Indicates the first The original electricity consumption values ​​of each collection point; This is the pre-processed power consumption value.

4. The intelligent load forecasting method integrating business logic and deep learning according to claim 1, characterized in that, The method for extracting the optimal typical day based on the business logic of similar day retrieval includes the following steps: S21, based on date hard constraint matching, select the historical day that is completely consistent with and closest to the date attribute of the prediction target day as the candidate reference day; S22, based on meteorological soft constraint matching, selects the historical day with the smallest deviation from the meteorological characteristics of the predicted target day among the candidate reference days as the optimal typical day.

5. The intelligent power load prediction method integrating business logic and deep learning according to claim 4, characterized in that, The method for calculating the deviation of meteorological characteristics between the candidate reference date and the predicted target date includes the following steps: S221, predicting the target date and the Meteorological feature matrix is ​​constructed from meteorological data collected at several electricity consumption data collection times for each candidate reference day: S2 22. Perform dimensionless processing on each eigenvalue in the meteorological feature matrix; S223, for each time step The predicted target date will be in this Meteorological characteristic sequence at time As a reference sequence, the first Meteorological characteristic sequence of candidate reference days As a comparison sequence, calculate each Correlation coefficient at time ; Indicates the predicted target date at time. The Dimensionless values ​​of meteorological characteristics; Indicates the first Candidate reference day at time The Dimensionless values ​​of meteorological characteristics; S224, calculate the... Each of the candidate reference days Correlation coefficient at time The mean of, as the first The comprehensive correlation between the meteorological characteristic sequences of the candidate reference days and the meteorological characteristic sequences of the predicted target days. and to Sort in descending order and filter out the items before sorting. The candidate reference date is selected as the optimal typical date.

6. The intelligent power load prediction method integrating business logic and deep learning according to claim 5, characterized in that, Correlation coefficient The calculation method is expressed as follows: in, Representing the The time of each candidate reference date The absolute difference of each meteorological feature; and These are the two minimum and two maximum differences across all times, all meteorological characteristics, and all candidate reference days; To take the minimum value among all candidate reference days; To take the minimum value among all meteorological characteristics; To find the minimum value over all times. Represents the resolution coefficient.

7. The intelligent power load prediction method integrating business logic and deep learning according to claim 1, characterized in that, In step S2, the method for determining the typical baseline electricity load curve for the target prediction day based on the optimal typical day is as follows: set up For the k-th optimal typical day, its 48-point load data is as follows: . like Predicting the daily electricity load baseline curve The electricity load data for the optimal typical day is directly taken; like Predicting the daily electricity load baseline curve The electricity load data for the two optimal typical days are generated by using either equal-weighted average or weighted average. The typical baseline electricity load curve uses the time of electricity consumption data collection as the horizontal axis and the electricity consumption data as the vertical axis.

8. The intelligent power load prediction method integrating business logic and deep learning according to claim 1, characterized in that, In step S3, the method for constructing the hybrid input of the model based on the multidimensional feature space is as follows: Historical time-series features, static covariant features, and business logic features carried in the multidimensional feature space are extracted, and then standardized and concatenated along the feature dimension. The concatenated result is used as the mixed input of the model. Historical time-series characteristics include historical load data for L consecutive days prior to the forecast target date; Static covariant features include date features and / or weather forecast features. Date features include one or more of the following: date attribute identifier, month, weekday, and time of the predicted target date. Weather forecast features include one or more of the following: temperature, humidity, cloud cover, and irradiance of the predicted target date. Business logic features include predicting the electricity load forecast curve for the target day; Feature standardization processing includes differential feature processing of extracted categorical and numerical features, including: For unordered category features, one-hot encoding is used to convert them into binary numerical vectors; For ordered category features, trigonometric function encoding is used to expand them into continuous periodic values; For continuous numerical features, Z-value standardization is used to transform continuous numerical values ​​into dimensionless standard values.

9. The intelligent power load prediction method integrating business logic and deep learning according to claim 1, characterized in that, Step S3, the method for constructing the neural network model includes the following steps: The constructed hybrid input and the measured electricity load data collected at various times on the current target day are used as training samples. The neural network model is trained through a BiLSTM bidirectional long short-term memory network. The training process is expressed as follows: in, and Represent The output of the LSTM layer is fed back from the previous and next time steps; represent Mixed feature input at time step; and These represent the weight matrices from the forward and backward LSTM layers to the output layer, respectively. Represents the bias term of the output layer; express The model output at time step; and Represent The output of the LSTM layer is fed back from the previous and next time steps; The above training process is supervised by the Huber loss function, and the model's gradient updates and weight configurations are based on the loss function values. When continuous... When the rate of decrease of the loss function value is less than the preset rate of decrease threshold, stop model training, save the model parameters, and obtain the target neural network model; The gradient update and weight configuration process calculates the model parameters for each training round. gradient of the loss function The model parameter weights are updated based on the gradient direction, and the calculation formula is as follows: in, These represent the parameters that the model needs to be trained, including the weight matrix and bias terms; Represents the learning rate; The value of the loss function; Indicates the first The parameters required for training the round model; Indicates the first The parameters required for training the round model.

Images