A power sale management system

Abstract

The application discloses a power sale management system capable of predicting power, and relates to the technical field of power sale management.The system comprises a data acquisition module for collecting basic data required for prediction, an input construction module for preprocessing and structure enhancing the basic data, and extracting a preliminary input feature sequence comprising a plurality of original input vectors from the basic data; and the input construction module generates corresponding mirror vectors according to preset symmetric structure construction rules, splices the mirror vectors with the original input vectors, and forms an extended input sequence.The application introduces mirror vectors and extended input sequences into the input construction module, significantly enhances the recognition ability of the model to trend mutation points and period jumps, and compared with an input mode relying only on original features, the method can establish symmetric feature expression at the input end, so that the prediction model has stronger structure perception ability when processing load reversal, seasonal fluctuations and the like.

Description

Technical Field

[0001] This invention relates to the field of electricity sales management technology, specifically to an electricity sales management system with predictable electricity consumption. Background Technology

[0002] In modern power systems, power forecasting is a crucial foundation for power dispatch and market transactions. Traditional power forecasting methods primarily rely on historical load data and basic environmental factors, using statistical regression models or time series models to achieve short-term or medium- to long-term predictions. These methods perform well when dealing with load scenarios that exhibit strong regularity and clear periodicity, supporting the stability of power system operation and the basic needs of the electricity market.

[0003] With the gradual increase in the proportion of renewable energy access and the diversification of user-side load characteristics, electricity forecasting faces more uncertainties. For example, extreme weather events caused by climate change, phased changes in user behavior, and dynamic adjustments to policies and electricity pricing mechanisms will all lead to new and complex characteristics in load data in terms of trends, fluctuations, and sudden disturbances.

[0004] Existing methods based on single time series or traditional regression analysis have certain limitations in predictive accuracy and adaptability when dealing with multidimensional perturbation conditions;

[0005] To improve the accuracy and robustness of electricity prediction, researchers have gradually introduced deep learning methods and multi-source data fusion approaches, attempting to use neural network models to capture complex nonlinear relationships. These methods can improve prediction performance to some extent, but two key issues still need to be addressed in long-term operation: first, how to enhance the structural expressive power of input features to better identify trend reversals and periodic jumps; and second, how to effectively model and respond to sudden disturbances to improve the adaptability of the prediction model in dynamic scenarios.

[0006] Therefore, proposing a power prediction system that takes into account both periodic correction and disturbance feedback mechanisms has strong practical significance and application value. Consequently, this application proposes a power sales management system that can predict power consumption. Summary of the Invention

[0007] The purpose of this invention is to provide a predictable electricity sales management system to solve the problems mentioned in the background section.

[0008] The present invention can be achieved through the following technical solution: a predictable electricity sales management system, comprising: a data acquisition module, an input construction module, an electricity prediction module, a period correction module, a disturbance feedback module, and an electricity sales plan generation module;

[0009] The data acquisition module is used to collect the basic data required for prediction. The basic data includes historical power data, environmental data, operating status data, contract information data, and scheduling event data.

[0010] The input construction module is used to preprocess and structurally enhance the basic data, including:

[0011] A preliminary input feature sequence is extracted from the basic data. The preliminary input feature sequence includes multiple feature vectors, each of which represents the input state under a time segment and is called the original input vector.

[0012] After normalizing and aligning the original input vector, the input construction module generates corresponding mirror vectors for preset key input features (such as temperature and load) according to preset symmetric structure construction rules. The mirror vectors are used to reflect the symmetric behavior of the input features under possible trend reversal conditions.

[0013] The input construction module concatenates the original input vector with the mirror vector to form an extended input sequence, enhancing the model's ability to identify trend abrupt changes.

[0014] The power prediction module is used to receive the extended input sequence and generate the first round of prediction results through a time-series neural network model, and outputs the first round of predicted power sequence within the prediction period.

[0015] The periodic correction module performs trend adjustments on the first round of predicted electricity sequence based on historical periodic error behavior;

[0016] The cycle correction module establishes multiple residual templates with fixed cycle lengths, such as 7-day, 14-day, and 30-day cycles. It constructs residual statistical curves for each type of cycle and matches the closest historical cycle template based on the time period label to which the current input feature belongs, thereby correcting the first round of prediction results and outputting the corrected predicted electricity sequence after cycle residual alignment.

[0017] The disturbance feedback module is used to compare the corrected predicted electricity consumption sequence output by the periodic correction module with the actual electricity consumption data of the corresponding time period, calculate the error between the two, and accumulate the error in a structured manner to form a disturbance residual chain. The disturbance feedback module extracts disturbance cause features through cluster analysis and pattern recognition methods, encodes the extracted disturbance causes into disturbance feature vectors, and embeds the disturbance feature vectors into the extended input sequence of the next round, so as to realize the feedforward absorption of disturbance sources by the prediction model and improve the robustness of the model to disturbance-prone scenarios.

[0018] The electricity sales plan generation module takes the periodically corrected predicted electricity volume sequence output by the disturbance feedback module as the final predicted electricity volume sequence, and matches it with the constraints of the electricity sales contract to generate plan information for electricity sales execution. The electricity sales plan includes predicted electricity volume, available sales boundaries, electricity price time period matching suggestions, performance risk warnings, etc., and may be used to directly drive the contract execution module or report to the power trading platform.

[0019] A further technical improvement of the present invention is that the step of the input construction module generating the corresponding mirror vector includes:

[0020] S1. Based on the feature importance assessment results in the basic data, select input variables that are highly correlated with load changes and have obvious periodic or trend reversal characteristics to form a key input feature set including multiple key input features;

[0021] S2. For each key input feature, determine its symmetric reference value based on its statistical characteristics within the training data or sliding time window. , serving as the symmetry center reference for mirror mapping;

[0022] The symmetric reference value is: the historical mean, the median of the distribution, or a specific threshold set by the user.

[0023] S3. For each original key input feature value Generate its corresponding mirror feature value according to the following formula. :

[0024] ;

[0025] S4. Combine the mirror values ​​of all key input features to form a mirror vector with the same dimension as the original key feature sub-vector. Then, concatenate the original input vector and the corresponding mirror vector along the feature dimension to obtain the extended input vector.

[0026] A further technical improvement of the present invention is that the method for constructing an extended input sequence using the input construction module includes:

[0027] Z1. Obtain the original input vector corresponding to the current prediction time. The original input vector consists of multiple original key input feature values, represented as follows:

[0028] ;

[0029] in, Let represent the i-th original key input feature value, and d be the number of feature dimensions;

[0030] Z2, for each original key input feature value Calculate the corresponding symmetric reference value based on its historical statistical attributes. And generate mirror feature values. ;

[0031] Z3, mirroring feature values ​​across all dimensions Arranged in sequence, they can be combined to form a mirror vector:

[0032] ;

[0033] Z4. Convert the original input vector With mirror vector The features are concatenated to form an expanded input vector:

[0034] ;

[0035] Z5. Within the set sliding window length L, extract the extended input vectors of continuous time steps to form the input sequence of the prediction model:

[0036] .

[0037] A further technical improvement of the present invention is that the step of outputting the corrected predicted energy sequence after correction by the periodic correction module includes:

[0038] Y1, The period correction module presets several fixed period lengths;

[0039] For each fixed period length, multiple historical period sample sequences are selected;

[0040] For each time point within each cycle, the difference between the actual electricity consumption and the predicted electricity consumption in the first round is calculated to form time series residual data;

[0041] The residual data are aligned according to their relative positions within the period to form a set of residuals for multiple historical periods;

[0042] Y2. For each type of cycle length, construct a residual statistical curve based on its historical cycle residual set;

[0043] Y3. The period correction module extracts the time period label corresponding to the prediction task based on the timestamp information in the input features at the current time.

[0044] Y4. Based on the time period label, select the historical period residual template that is most similar to the current prediction input from the residual statistical curve;

[0045] Y5. Correct the residual values ​​at each time point in the selected period residual template by matching them one-to-one with the first round of prediction results:

[0046] Corrected forecast value = First round forecast value + Residual value at the corresponding time point;

[0047] Y6. Output the corrected predicted power sequence after the periodic residual alignment as the output of the periodic correction module.

[0048] A further technical improvement of the present invention is that: the electricity sales plan generation module divides the total electricity sales contract fulfillment volume into multiple fine-grained segments, and dynamically determines whether each segment should enter the execution state based on the prediction deviation, thereby realizing prediction-driven segment-by-segment contract fulfillment control, including:

[0049] A1. The performance period of the target electricity sales contract According to the preset time granularity Divided into N= / Given equal time intervals, we obtain a contract fragment set C = { , ,..., };in, = ( , ), representing the i-th contract segment with an execution period. and target electricity delivery ;

[0050] A2. Obtain segment prediction values ​​by extracting the predicted values ​​for each segment from the final predicted power sequence output by the disturbance feedback module. Corresponding predicted value ,Right now ;in, This indicates the prediction model for the time period. The predicted power consumption value;

[0051] A3. The system calculates the prediction error for each contract segment: At the same time, set a deviation threshold. ;

[0052] A4. For each segment :

[0053] like ≤ If so, the segment is marked as "executable";

[0054] like > If the fragment is frozen, it will not be executed and will be pushed into the waiting queue Q.

[0055] The frozen fragment is recorded as: Q={ | > };

[0056] A5. The system continues to re-predict in subsequent time periods;

[0057] Whenever a new prediction is updated, the fragments in the frozen fragment queue Q are recalculated. And make a judgment:

[0058] If satisfied ≤ If so, the fragment will be unfrozen and added to the most recent executable cycle for realization;

[0059] If the deviation requirements are not met and the maximum extension window is exceeded, the process will proceed to manual scheduling or default handling.

[0060] A6. During settlement, the actual electricity consumption shall be calculated on a per-contract basis. Compliance with target electricity volume Perform deviation calculation and record the deviation rate. :

[0061] ;

[0062] Furthermore, the system is based on the deviation rate of each executed contract segment. The pre-defined reward and punishment segmentation strategy is used to implement tiered reward and punishment settlements.

[0063] A further technical improvement of the present invention is that: the power prediction module detects trend change points in the extended input sequence, and selects a target prediction model from multiple preset candidate prediction models based on the current input feature state to perform power prediction;

[0064] The criteria for judging trend abrupt change points include whether the rate of change of key input features exceeds a preset slope threshold, or whether the residual standard deviation exceeds a preset residual abrupt change threshold.

[0065] Specifically, including:

[0066] H1. Real-time reception of the extended input sequence output by the input construction module. , This represents the extended input vector at the current time t;

[0067] And set the sliding window length W, and calculate the slope, gradient, and residual standard deviation of the key input feature dimensions at continuous time steps;

[0068] H2. When the rate of change of a key input feature dimension exceeds a preset slope threshold, or the residual standard deviation exceeds a residual mutation threshold, the system determines that the current input sequence has a "trend bifurcation" and generates a bifurcation flag. ;

[0069] H3. The system maintains multiple candidate prediction models, and when the bifurcation sign... At that time, the system evaluates each candidate model by matching it with a similarity scoring function based on the context state of the current input features, and selects the target prediction model that is most similar to the current state for prediction.

[0070] H4. After the model reconstruction is completed, the predicted electricity sequence is generated. Where H is the prediction step size;

[0071] At the same time, the bifurcation marker The selected model number and switching timestamp information are transmitted to the disturbance feedback module.

[0072] The system will output the extended input sequence from the input construction module. As input to the target prediction model; where each It contains a vector composed of the original key input features and the mirrored features, with a dimension of 2d;

[0073] The target prediction model receives an extended input sequence. Then, a complete forward propagation operation is performed, and the predicted power consumption value within the future prediction time window is recursively output through the multi-layer structure of the neural network.

[0074] The target prediction model outputs a power prediction sequence within the prediction time range [t+1, t+H], expressed as:

[0075] ;in, This represents the predicted power consumption value for the h-th time step in the future.

[0076] Compared with the prior art, the present invention has the following beneficial effects:

[0077] This invention significantly enhances the model's ability to identify trend abrupt changes and periodic jumps by introducing mirror vectors and extended input sequences into the input construction module. Compared with input methods that rely solely on original features, this method can establish symmetrical feature representations at the input end, enabling the prediction model to have stronger structural awareness when dealing with scenarios such as load reversal and seasonal fluctuations, thereby effectively improving the stability and accuracy of prediction results.

[0078] Furthermore, this invention establishes multiple periodic residual templates through a periodic correction module and dynamically matches them with the currently input time period label, thereby achieving periodic correction of prediction errors. This mechanism can fully utilize historical periodic patterns, align and compensate residual characteristics with predicted values, and improve prediction results in terms of periodic consistency and error stability, providing more valuable prediction data for the medium- and long-term operation of power systems.

[0079] On the other hand, this invention uses a disturbance feedback module to perform cluster analysis and pattern recognition on external disturbance events, generating disturbance feature vectors and embedding them into the next round of prediction. This enables structured absorption and feedforward adjustment of disturbance information, allowing the prediction system to have higher adaptability when dealing with scenarios such as electricity price fluctuations, extreme weather, or sudden load events. At the same time, the system combines fragmented performance of electricity sales contracts with a reward and punishment mechanism, providing reliable data support for the generation of electricity sales plans, which helps to improve the performance rate and economy of electricity transactions. Attached Figure Description

[0080] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0081] Figure 1 This is a schematic diagram of the system logic of the present invention. Detailed Implementation

[0082] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0083] Example 1

[0084] Please see Figure 1 As shown, the present invention provides a predictable electricity sales management system, including: a data acquisition module, an input construction module, an electricity prediction module, a period correction module, a disturbance feedback module, and an electricity sales plan generation module;

[0085] The data acquisition module is used to collect the basic data required for forecasting. The basic data includes historical electricity data, environmental data, operating status data, contract information data, and scheduling event data.

[0086] Among them, historical electricity data is the actual electricity consumption record of the target object in multiple periods;

[0087] Environmental data include external influencing factors that are significantly related to load changes, such as temperature, humidity, wind speed, and light intensity;

[0088] Operating status data includes the device's start / stop status, voltage, current, etc.

[0089] Contract information data includes the contracted electricity volume range, electricity price structure, and execution period in the electricity sales contract;

[0090] Dispatch event data includes control instructions, load switching commands, etc.

[0091] The input construction module is used to preprocess and structure the underlying data, including:

[0092] A preliminary input feature sequence is extracted from the basic data. The preliminary input feature sequence includes multiple feature vectors, each of which represents the input state in a time segment and is called the original input vector.

[0093] Specifically, the original input vector may consist of one or more of the following:

[0094] User unit power consumption within the current time segment;

[0095] The average ambient temperature value within the current time segment;

[0096] Electricity price level in the current time segment;

[0097] Is the current time segment during a holiday, workday, or a special electricity control period?

[0098] The weather category code corresponding to the current time segment (such as sunny, rainy, cloudy, snowy, etc.);

[0099] Does the current time segment contain any recorded disturbance events?

[0100] The rate of change in electricity volume compared to the previous time segment;

[0101] Peak-valley electricity price category identification within the current time segment;

[0102] User type coding (e.g., industrial user, commercial user, residential user);

[0103] Other key parameters related to system operation (such as energy storage charge and discharge status, predicted power supply, historical prediction error, etc.).

[0104] The original input vector is constructed by encoding the above-mentioned information into numerical or discrete features, forming a fixed-length, multi-dimensional vector representation. The number of dimensions depends on the types of features selected. The original input vectors generated from all time segments are arranged in chronological order to form a preliminary input feature sequence, which serves as the basic input data for subsequent prediction input construction modules.

[0105] After normalizing and aligning the original input vector, the input construction module generates corresponding mirror vectors for preset key input features (such as temperature and load) according to preset symmetric structure construction rules. The mirror vectors are used to reflect the symmetric behavior of the input features under possible trend reversal conditions.

[0106] Key input features refer to characteristic variables that are significantly correlated with load fluctuation trends and have a structural impact on electricity sales forecasting during the change of electricity sales volume of the target object. Preferably, key input features include at least periodic or trend-based external factors such as temperature, load, solar radiation intensity, and wind speed, specifically:

[0107] Temperature characteristics are used to reflect the driving behavior of the environment on temperature-controlled loads such as air conditioning and electric heating;

[0108] Load characteristics are used to capture changes in the electrical intensity of a target object;

[0109] Sunlight intensity and wind speed can indirectly reflect the access of new energy sources or the regulation of heating and cooling loads.

[0110] The steps for generating the corresponding mirror vector from the input constructor module include:

[0111] S1. Determine the key input feature set:

[0112] Based on the feature importance assessment results in the basic data, input variables that are highly correlated with load changes and have obvious periodic or trend reversal characteristics are selected to form a key input feature set including multiple key input features;

[0113] S2. Construct a symmetric reference value:

[0114] For each key input feature, a symmetric reference value is determined based on its statistical properties within the training data or a sliding time window. The symmetric reference values ​​are: historical mean, median of distribution, or a specific threshold set by the individual.

[0115] The symmetry reference value is used as the symmetry center reference for mirror mapping;

[0116] S3. Perform symmetric mapping operation:

[0117] For each original key input feature value Generate its corresponding mirror feature value according to the following formula. :

[0118] ;

[0119] S4. Construct and concatenate mirror vectors:

[0120] The mirror values ​​of all key input features are combined to form a mirror vector with the same dimension as the original key feature sub-vectors. Then, the original input vector and the corresponding mirror vector are concatenated along the feature dimension to obtain the extended input vector.

[0121] The input construction module concatenates the original input vector with the mirror vector to form an extended input sequence, enhancing the model's ability to identify trend abrupt changes.

[0122] Methods for constructing extended input sequences using input building blocks include:

[0123] Z1. Obtain the original input vector:

[0124] Obtain the original input vector corresponding to the current prediction time. The original input vector consists of multiple original key input feature values, represented as follows:

[0125] ;

[0126] in, Let represent the i-th original key input feature value, and d be the number of feature dimensions;

[0127] Z2. Generate the corresponding mirror feature values:

[0128] For each original key input feature value Calculate the corresponding symmetric reference value based on its historical statistical attributes. And generate mirror feature values. ;

[0129] Z3. Construct a mirror vector:

[0130] Mirrored feature values ​​in all dimensions Arranged in sequence, they can be combined to form a mirror vector:

[0131] ;

[0132] Z4. Concatenate the original vector and its mirror image to generate an expanded input vector:

[0133] The original input vector With mirror vector The features are concatenated to form an expanded input vector:

[0134] ;

[0135] Z5. Constructing the extended input sequence:

[0136] Within a set sliding window length L, the extended input vectors of consecutive time steps are extracted to form the input sequence of the prediction model:

[0137] ;

[0138] Through the above steps, the input construction module can construct an extended input sequence that simultaneously contains the original key input feature values ​​and their corresponding mirror feature values, enabling the prediction model to have a stronger structural awareness capability when dealing with scenarios such as load trend reversal, periodic jumps, or abnormal disturbances.

[0139] The power prediction module is used to receive the extended input sequence and generate the first round of prediction results through a time-series neural network model. This model can be a Transformer network, a Long Short-Term Memory network (LSTM) or a combination thereof, and the output is the first round of predicted power sequence within the prediction period.

[0140] The periodic correction module performs trend adjustments on the first round of predicted electricity volume sequence based on historical periodic error behavior;

[0141] The cycle correction module establishes multiple residual templates with fixed cycle lengths, such as 7-day, 14-day, and 30-day cycles. It constructs residual statistical curves for each type of cycle and matches the closest historical cycle template based on the time period label to which the current input feature belongs, thereby correcting the first round of prediction results and outputting the corrected predicted electricity sequence after cycle residual alignment.

[0142] The steps for correcting the predicted energy sequence after the periodic correction module outputs the corrected data include:

[0143] Y1. Establish multiple residual templates with fixed period lengths:

[0144] The cycle correction module presets several fixed cycle lengths, preferably including 7-day, 14-day and 30-day cycles, which correspond to common power fluctuation patterns such as weekly cycle, half-month cycle and monthly cycle, respectively.

[0145] For each fixed period length, multiple historical period sample sequences are selected;

[0146] For each time point within each cycle (such as every hour or every hour of the day), calculate the difference between the actual electricity consumption and the electricity consumption predicted in the first round to form time series residual data;

[0147] The residual data are aligned according to their relative positions within the period to form a set of residuals for multiple historical periods;

[0148] Y2. Constructing the periodic residual statistical curve:

[0149] For each type of cycle length, a residual statistical curve is constructed based on its historical cycle residual set. This residual statistical curve reflects the changing trend of typical residual values ​​at each time point within the cycle and is used for subsequent prediction correction.

[0150] Residual statistics curves can be constructed in the following ways:

[0151] Calculate the average residual (or median, weighted mean) at each time point within the calculation period.

[0152] A standard residual template with a length consistent with that period is formed.

[0153] Each residual statistical curve includes a label for that period, which is used for matching.

[0154] Y3. Determine the period label of the current prediction input:

[0155] The period correction module extracts the time period label corresponding to the prediction task based on the timestamp information (such as day of the week, whether it is a working day, date position, etc.) in the input features at the current time.

[0156] The current cycle type can be determined by the "time location information" field in the feature vector, such as "week index" and "month index";

[0157] This time period label is used to subsequently match the most similar historical period template;

[0158] Y4. Match the closest historical period residual template:

[0159] The periodic correction module selects the template that best matches the current forecast period label from the existing residual statistical curves:

[0160] A time-based matching method can be used (e.g., if the current time is Tuesday, select a 7-day cycle template).

[0161] Similarity calculation rules can also be introduced to calculate the distance between the current input feature (such as load level, weather temperature, etc.) and the mean of the input features of each template in the historical period, so as to select the most similar residual template;

[0162] Y5. Revise the results of the first round of predictions:

[0163] The residual values ​​at each time point in the selected period residual template are matched one-to-one with the first round of prediction results for correction.

[0164] The correction method is as follows:

[0165] Corrected forecast value = First-round forecast value + Residual value at the corresponding time point

[0166] Maintaining the consistency of the cyclical structure helps to enhance the cyclical consistency and stability of the forecast results;

[0167] Y6. Corrected predicted energy sequence after output periodic residual alignment:

[0168] After the periodic correction module completes the above residual compensation, it outputs a set of corrected predicted power sequence with the same length as the first round of prediction results. This sequence has higher periodic consistency and residual fitting degree, and is passed to subsequent modules or user systems as the final output result.

[0169] The disturbance feedback module compares the corrected predicted electricity consumption sequence output by the periodic correction module with the actual electricity consumption data for the corresponding period, calculates the error between the two, and accumulates the error in a structured manner to form a disturbance residual chain. The disturbance feedback module extracts disturbance cause features through cluster analysis and pattern recognition methods, encodes the extracted disturbance causes into disturbance feature vectors, and embeds the disturbance feature vectors into the extended input sequence of the next round, so as to realize the feedforward absorption of disturbance sources by the prediction model and improve the robustness of the model to disturbance-prone scenarios.

[0170] In this embodiment, the step of the perturbation feedback module embedding the perturbation feature vector into the extended input sequence of the next round includes:

[0171] First, based on a multi-source heterogeneous dataset consisting of historical electricity consumption sequences of the target region and related external events (such as policy adjustments, electricity price fluctuations, extreme weather, and sudden load events), clustering analysis is used to classify time segments with similar disturbance characteristics. Preferably, the clustering method may include, but is not limited to, K-means clustering, DBSCAN density clustering, or soft clustering methods based on Gaussian mixture models, to identify the distribution patterns and similar regions of disturbance events in the time series.

[0172] After classifying the disturbance events, the system further uses pattern recognition algorithms (such as support vector machines, convolutional neural networks, and temporal feature encoders) to extract features and encode vectors for each disturbance type, generating a unified representation of the disturbance cause. This representation is defined as a disturbance feature vector. This disturbance feature vector contains information representing dimensions such as disturbance category, disturbance duration, intensity level, and historical response impact, forming a numerical vector of uniform length, which serves as the standard output of the disturbance feedback module.

[0173] Next, when the prediction model executes the next prediction iteration, the perturbation feedback module embeds the perturbation feature vector as external feedback information into the extended input sequence of the next round. Specifically, the perturbation feature vector is concatenated with the mirror vector and periodic correction vector corresponding to the current time step in the extended input sequence to form a multi-dimensional extended input vector that integrates perturbation information, which participates in the subsequent time series prediction modeling process.

[0174] Through the above methods, the disturbance feedback module can realize the feature learning and structured expression of historical disturbance events, and can continuously update the feedback to the input end to help the prediction model dynamically adjust its sensitivity to disturbances, thereby effectively improving the accuracy and responsiveness of power prediction and meeting the intelligent prediction needs in complex power consumption scenarios.

[0175] The electricity sales plan generation module is used to take the periodically corrected predicted electricity sequence output by the disturbance feedback module as the final predicted electricity sequence, and match it with the electricity sales contract constraints to generate plan information for electricity sales execution.

[0176] The electricity sales plan includes forecasted electricity volume, available sales boundaries, electricity price time period matching suggestions, performance risk warnings, etc., and may be used to directly drive the contract execution module or report to the electricity trading platform.

[0177] Furthermore, the power prediction module detects trend abrupt changes in the extended input sequence and selects a target prediction model from multiple preset candidate prediction models based on the current input feature state to predict power consumption.

[0178] The criteria for judging trend abrupt change points include whether the rate of change of key input features exceeds a preset slope threshold, or whether the residual standard deviation exceeds a preset residual abrupt change threshold.

[0179] Specifically, including:

[0180] H1, Input Sequence Monitoring:

[0181] Real-time reception of extended input sequences output by the input construction module , This represents the extended input vector at the current time t;

[0182] And set the sliding window length W, and calculate the slope, gradient, and residual standard deviation of the key input feature dimensions at continuous time steps;

[0183] H2. Bifurcation point trigger condition judgment:

[0184] When the system detects that the rate of change of a key input feature dimension exceeds a preset slope threshold, or that the residual standard deviation exceeds a residual mutation threshold, it determines that the current input sequence has experienced a "trend bifurcation" and generates a bifurcation flag. ;

[0185] H3. Dynamic Model Path Selection:

[0186] The system maintains multiple candidate prediction models, and when the bifurcation flag... At that time, the system evaluates each candidate model by matching it with a similarity scoring function based on the context state of the current input features, and selects the target prediction model that is most similar to the current state for prediction.

[0187] The similarity score is constructed based on the fitting error performance of each model in the historical training set under similar input feature conditions, and is used to dynamically switch or activate the corresponding model;

[0188] H4. Predicted Output and Feedback:

[0189] After the model reconstruction is completed, a predicted electricity sequence is generated. Where H is the prediction step size;

[0190] At the same time, the bifurcation marker The selected model number and switching timestamp information are transmitted to the disturbance feedback module for archiving, storage and causal analysis to support subsequent disturbance cause extraction and feedforward adjustment.

[0191] Specifically, after receiving the model switching control command, the system loads the corresponding candidate prediction model from the model library based on the selected model identifier. Candidate prediction model It may include temporal deep neural network structures built based on Transformer, LSTM, GRU or a combination thereof, which have been trained with historical data and stored in the model call interface in a modular manner;

[0192] The system will output the extended input sequence from the input construction module. As input to the target prediction model; where each It contains a vector composed of the original key input features and the mirrored features, with a dimension of 2d;

[0193] The target prediction model receives an extended input sequence. Then, a complete forward propagation operation is performed, and the predicted power consumption value within the future prediction time window is recursively output through the multi-layer structure of the neural network.

[0194] The target prediction model outputs a power prediction sequence within the prediction time range [t+1, t+H], expressed as:

[0195] ;in, This represents the predicted power consumption value for the h-th time step in the future.

[0196] Example 2

[0197] A predictable electricity sales management system includes: a data acquisition module, an input construction module, an electricity prediction module, a period correction module, a disturbance feedback module, and an electricity sales plan generation module.

[0198] The data acquisition module is used to collect the basic data required for forecasting. The basic data includes historical electricity data, environmental data, operating status data, contract information data, and scheduling event data.

[0199] The input construction module is used to preprocess and structure the underlying data, including:

[0200] A preliminary input feature sequence is extracted from the basic data. The preliminary input feature sequence includes multiple feature vectors, each of which represents the input state in a time segment and is called the original input vector.

[0201] After normalizing and aligning the original input vector, the input construction module generates corresponding mirror vectors for preset key input features (such as temperature and load) according to preset symmetric structure construction rules. The mirror vectors are used to reflect the symmetric behavior of the input features under possible trend reversal conditions.

[0202] The input construction module concatenates the original input vector with the mirror vector to form an extended input sequence, enhancing the model's ability to identify abrupt trend changes.

[0203] The power prediction module is used to receive the extended input sequence and generate the first round of prediction results through a time-series neural network model. This model can be a Transformer network, a Long Short-Term Memory network (LSTM) or a combination thereof, and the output is the first round of predicted power sequence within the prediction period.

[0204] The periodic correction module performs trend adjustments on the first round of predicted electricity volume sequence based on historical periodic error behavior;

[0205] The cycle correction module establishes multiple residual templates with fixed cycle lengths, such as 7-day, 14-day, and 30-day cycles. It constructs residual statistical curves for each type of cycle and matches the closest historical cycle template based on the time period label to which the current input feature belongs, thereby correcting the first round of prediction results and outputting the corrected predicted electricity sequence after cycle residual alignment.

[0206] The disturbance feedback module compares the corrected predicted electricity consumption sequence output by the periodic correction module with the actual electricity consumption data for the corresponding period, calculates the error between the two, and accumulates the error in a structured manner to form a disturbance residual chain. The disturbance feedback module extracts disturbance cause features through cluster analysis and pattern recognition methods, encodes the extracted disturbance causes into disturbance feature vectors, and embeds the disturbance feature vectors into the extended input sequence of the next round, so as to realize the feedforward absorption of disturbance sources by the prediction model and improve the robustness of the model to disturbance-prone scenarios.

[0207] The electricity sales plan generation module takes the periodically corrected predicted electricity volume sequence output by the disturbance feedback module as the final predicted electricity volume sequence, matches it with the constraints of the electricity sales contract, and generates plan information for electricity sales execution. The electricity sales plan includes predicted electricity volume, available sales boundaries, electricity price time period matching suggestions, performance risk warnings, etc., and can be used to directly drive the contract execution module or report to the power trading platform.

[0208] Compared to Example 1, the electricity sales plan generation module in Example 2 divides the total electricity sales contract fulfillment volume into multiple fine-grained segments, and dynamically determines whether each segment enters the execution state based on the prediction deviation, thereby achieving prediction-driven segment-by-segment contract fulfillment control, including:

[0209] A1. Contract Segment Division:

[0210] The performance period of the target electricity sales contract According to the preset time granularity Divided into N= / Given equal time intervals, we obtain a contract fragment set C = { , ,..., };in, = ( , ), representing the i-th contract segment with an execution period. and target electricity delivery ;

[0211] A2. Obtain segment prediction values ​​by extracting the predicted values ​​for each segment from the final predicted power sequence output by the disturbance feedback module. Corresponding predicted value ,Right now ;in, This indicates the prediction model for the time period. The predicted power consumption value;

[0212] A3. Calculate the prediction deviation:

[0213] The system calculates the prediction error for each contract segment: At the same time, set a deviation threshold. This is used to determine whether a fragment meets the execution conditions.

[0214] A4. Fragment Status Determination and Freezing Mechanism:

[0215] For each segment :

[0216] like ≤ If so, the segment is marked as "executable";

[0217] like > If the fragment is frozen, it will not be executed and will be pushed into the waiting queue Q.

[0218] The frozen fragment is recorded as: Q={ | > };

[0219] A5. Deferred Redemption and Recovery Mechanism:

[0220] The system continues to re-predict in subsequent time periods;

[0221] Whenever a new prediction is updated, the fragments in the frozen fragment queue Q are recalculated. And make a judgment:

[0222] If satisfied ≤ If so, the fragment will be unfrozen and added to the most recent executable cycle for realization;

[0223] If the deviation requirements are not met and the maximum extension window is exceeded, the process will proceed to manual scheduling or default handling.

[0224] A6. Independent Segment Settlement Mechanism:

[0225] During settlement, the actual electricity consumption will be calculated on a per-contract basis. Compliance with target electricity volume Perform deviation calculation and record the deviation rate. :

[0226] ;

[0227] Furthermore, the system is based on the deviation rate of each executed contract segment. In this embodiment, tiered reward and punishment settlements are performed separately from the preset reward and punishment segmentation strategy. Specifically:

[0228] If the deviation rate is lower than the first judgment threshold, a weighted reward will be applied;

[0229] If the price falls within the middle range, the settlement will be based on the standard contract price.

[0230] If the deviation rate exceeds the maximum tolerance limit, a penalty will be imposed on the segment, or it will be included in the risk assessment indicators.

[0231] The reward and punishment rules can be set by the trading platform or customized by the user.

[0232] For example:

[0233] If the percentage is ≤5%, the reward will be calculated at 102%.

[0234] 5% < ≤10%, settlement will be based on the contract price;

[0235] >10%, penalty for breach of contract.

[0236] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0237] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A predictable electricity sales management system, characterized in that, include: The data acquisition module is used to collect the basic data required for prediction; The input construction module is used to preprocess and structurally enhance the basic data, and extract a preliminary input feature sequence containing multiple original input vectors. Furthermore, the input construction module generates corresponding mirror vectors for preset key input features based on preset symmetric structure construction rules, and concatenates the mirror vectors with the original input vectors to form an extended input sequence; The power prediction module generates the first round of prediction results based on the extended input sequence through a time-series neural network model, and outputs the first round of predicted power sequence within the prediction period. The periodic correction module performs trend adjustments on the first round of predicted electricity generation sequence based on historical periodic error behavior and outputs the corrected predicted electricity generation sequence. The perturbation feedback module is used to compare the corrected predicted electricity consumption sequence with the actual electricity consumption data of the corresponding time period to form a perturbation residual chain; and to extract the perturbation cause features through cluster analysis and pattern recognition methods, encode them into perturbation feature vectors, and embed them into the next round of extended input sequence. The electricity sales plan generation module matches the predicted electricity volume sequence output by the disturbance feedback module with the constraints of the electricity sales contract to generate electricity sales execution plan information; The step of generating the corresponding mirror vector by the input construction module includes: S1. Based on the feature importance assessment results in the basic data, select input variables that are highly correlated with load changes and have obvious periodic or trend reversal characteristics to form a key input feature set including multiple key input features; S2. For each key input feature, determine its symmetric reference value based on its statistical characteristics within the training data or sliding time window. , serving as the symmetry center reference for mirror mapping; The symmetric reference value is: the historical mean, the median of the distribution, or a specific threshold set by the user. S3. For each original key input feature value Generate its corresponding mirror feature value according to the following formula. : ； S4. Combine the mirror values ​​of all key input features to form a mirror vector with the same dimension as the original key feature sub-vector. Then, concatenate the original input vector with the corresponding mirror vector along the feature dimension to obtain the extended input vector. The method for constructing an extended input sequence using the input construction module includes: Z1. Obtain the original input vector corresponding to the current prediction time. The original input vector consists of multiple original key input feature values, represented as follows: ； in, Let represent the i-th original key input feature value, and d be the number of feature dimensions; Z2, for each original key input feature value Calculate the corresponding symmetric reference value based on its historical statistical attributes. And generate mirror feature values. ; Z3, mirroring feature values ​​across all dimensions Arranged in sequence, they can be combined to form a mirror vector: ； Z4. Convert the original input vector With mirror vector The features are concatenated to form an expanded input vector: ； Z5. Within the set sliding window length L, extract the extended input vectors of continuous time steps to form the input sequence of the prediction model: ； The step of the periodic correction module outputting the corrected predicted energy sequence includes: Y1, The period correction module presets several fixed period lengths; For each fixed period length, multiple historical period sample sequences are selected; For each time point within each cycle, the difference between the actual electricity consumption and the predicted electricity consumption in the first round is calculated to form time series residual data; The residual data are aligned according to their relative positions within the period to form a set of residuals for multiple historical periods; Y2. For each type of cycle length, construct a residual statistical curve based on its historical cycle residual set; Y3. The period correction module extracts the time period label corresponding to the prediction task based on the timestamp information in the input features at the current time. Y4. Based on the time period label, select the historical period residual template that is most similar to the current prediction input from the residual statistical curve; Y5. Correct the residual values ​​at each time point in the selected period residual template by matching them one-to-one with the first round of prediction results: Corrected forecast value = First round forecast value + Residual value at the corresponding time point; Y6. Output the corrected predicted power sequence after the periodic residual alignment as the output of the periodic correction module; The electricity sales plan generation module divides the total electricity sales contract into multiple fine-grained segments and dynamically determines whether each segment should enter the execution state based on the prediction deviation of each segment. The method for the electricity sales plan generation module to determine the dynamics of prediction deviation includes: A1. The performance period of the target electricity sales contract According to the preset time granularity Divided into N= / Given equal time segments, we obtain the contract segment set C = { , ,..., };in, = ( , ), representing the i-th contract segment with an execution period. and target electricity delivery ; A2. Obtain segment prediction values ​​by extracting the predicted values ​​for each segment from the final predicted power sequence output by the disturbance feedback module. Corresponding predicted value ,Right now ;in, This indicates the prediction model for the time period. The predicted power consumption value; A3. The system calculates the prediction error for each contract segment: At the same time, set a deviation threshold. ; A4. For each segment : like ≤ If so, the segment is marked as "executable"; like > If the fragment is frozen, it will not be executed and will be pushed into the waiting queue Q. The frozen fragment is recorded as: Q={ | > }; A5. The system continues to re-predict in subsequent time periods; Whenever a new prediction is updated, the fragments in the frozen fragment queue Q are recalculated. And make a judgment: If satisfied ≤ If so, the fragment will be unfrozen and added to the most recent executable cycle for realization; If the deviation requirements are not met and the maximum extension window is exceeded, the process will proceed to manual scheduling or default handling. A6. During settlement, the actual electricity consumption shall be calculated on a per-contract basis. Compliance with target electricity volume Perform deviation calculation and record the deviation rate. : ； Furthermore, the system is based on the deviation rate of each executed contract segment. In contrast to the preset reward and punishment segmentation strategy, tiered reward and punishment settlements are executed separately. The power prediction module detects trend abrupt changes in the extended input sequence and selects a target prediction model from multiple preset candidate prediction models based on the current input feature state to predict power consumption, including: H1. Real-time reception of the extended input sequence output by the input construction module. , This represents the extended input vector at the current time t; And set the sliding window length W, and calculate the slope, gradient, and residual standard deviation of the key input feature dimensions at continuous time steps; H2. When the rate of change of a key input feature dimension exceeds a preset slope threshold, or the standard deviation of the residual exceeds a residual mutation threshold, the system determines that the current input sequence has a "trend bifurcation" and generates a bifurcation flag. ; H3. The system maintains multiple candidate prediction models, and when the bifurcation sign... At that time, the system evaluates each candidate model by matching it with a similarity scoring function based on the context state of the current input features, and selects the target prediction model that is most similar to the current state for prediction. H4. After the model reconstruction is completed, the predicted electricity sequence is generated. Where H is the prediction step size; At the same time, the bifurcation marker The selected model number and switching timestamp information are transmitted to the disturbance feedback module.

2. The electricity sales management system with predictable electricity consumption according to claim 1, characterized in that, The specific steps of H4 include: After receiving the model switching control command, the system loads the corresponding candidate prediction model from the model library according to the selected model identifier. ; The system will output the extended input sequence from the input construction module. As input to the target prediction model; The target prediction model receives an extended input sequence. Then, a complete forward propagation operation is performed, and the predicted power consumption value within the future prediction time window is recursively output through the multi-layer structure of the neural network. The target prediction model outputs a power prediction sequence within the prediction time range [t+1, t+H], expressed as: ;in, This represents the predicted power consumption value for the h-th time step in the future.

Images