CN-121119767-B - Electricity selling management system capable of predicting electric quantity

Abstract

The invention discloses an electricity selling management system capable of predicting electric quantity, which relates to the technical field of electricity selling management and comprises a data acquisition module, an input construction module and an expansion input sequence, wherein the data acquisition module is used for acquiring basic data required by prediction, the input construction module is used for preprocessing and structurally reinforcing the basic data and extracting a preliminary input characteristic sequence containing a plurality of original input vectors from the basic data, and the input construction module generates corresponding mirror image vectors for preset key input characteristics according to preset symmetrical structure construction rules and splices the mirror image vectors with the original input vectors to form the expansion input sequence. According to the method, the mirror image vector and the extended input sequence are introduced into the input construction module, so that the recognition capability of the model to trend mutation points and cycle jumps is remarkably enhanced, and compared with an input mode which only depends on original features, the method can establish symmetrical feature expression at the input end, so that the prediction model has stronger structural perceptibility when processing scenes such as load reversal and seasonal fluctuation.

Inventors

• GUO JIANYUN

Assignees

• 广东惠电佳能源有限公司

Dates

Publication Date

20260512

Application Date

20250905

Claims (2)

1. 1. An electricity vending management system capable of predicting electricity quantity, comprising: The data acquisition module is used for acquiring basic data required by prediction; the input construction module is used for preprocessing and structurally enhancing the basic data and extracting a preliminary input characteristic sequence containing a plurality of original input vectors from the basic data; The input construction module generates a corresponding mirror image vector according to a preset symmetrical structure construction rule aiming at preset key input characteristics, and splices the mirror image vector with an original input vector to form an extended input sequence; the electric quantity prediction module is used for generating a first round of prediction result through the time sequence neural network model based on the extended input sequence and outputting a first round of predicted electric quantity sequence in a prediction period; The cycle correction module is used for carrying out trend adjustment on the first-round predicted electric quantity sequence based on the historical periodic error behavior and outputting a corrected predicted electric quantity sequence; The disturbance feedback module is used for comparing the corrected predicted electric quantity sequence with actual electric quantity data of a corresponding period to form a disturbance residual error chain, extracting disturbance inducement characteristics through a cluster analysis and pattern recognition method, encoding the disturbance inducement characteristics into disturbance characteristic vectors, and embedding the disturbance characteristic vectors into a next round of expansion input sequence; the electricity selling plan generating module is used for matching the predicted electric quantity sequence output by the disturbance feedback module with the constraint condition of the electricity selling contract to generate electricity selling execution plan information; the step of generating the corresponding mirror image vector by the input construction module comprises the following steps: S1, selecting an input variable which is highly related to load change and has obvious periodicity or trend reversal characteristics according to a characteristic importance evaluation result in the basic data to form a key input characteristic set comprising a plurality of key input characteristics; s2, determining symmetrical reference values of each key input feature based on statistical characteristics of the key input features in training data or sliding time window As a symmetry center reference for the mirror image map; the symmetrical reference value is a history mean value, a distribution median value or a specific threshold value set by people; S3, inputting characteristic values to each original key The corresponding mirror characteristic value is generated according to the following formula : ; S4, combining mirror image values of all key input features to form mirror image vectors, wherein the dimensions of the mirror image vectors are consistent with the original key feature sub-vectors, and then splicing the original input vectors and the corresponding mirror image vectors in feature dimensions to obtain extended input vectors; the method for forming the extended input sequence by the input construction module comprises the following steps: Z1, acquiring an original input vector corresponding to the current prediction moment, wherein the original input vector consists of a plurality of original key input characteristic values and is expressed as follows: ; Wherein, the Representing an ith original key input characteristic value, wherein d is a characteristic dimension number; Z2, for each original key input feature value Calculating corresponding symmetrical reference values based on historical statistical attributes thereof And generate mirror image characteristic value ; Z3, mirror image characteristic value under all dimensions Sequentially arranged, and combined to form a mirror image vector: ; Z4, the original input vector And mirror vector Splicing in the characteristic dimension to form an expansion input vector: ; Z5, extracting extended input vectors of continuous time steps within the set sliding window length L to form an input sequence of a prediction model: ; The step of outputting the corrected predicted electric quantity sequence by the period correction module comprises the following steps: y1, presetting a plurality of fixed cycle lengths by a cycle correction module; for each fixed cycle length, selecting a plurality of historical cycle sample sequences; Calculating the difference value between the real electric quantity and the first-round predicted electric quantity at each time point in each period to form time sequence residual error data; aligning residual data according to relative positions in the period to form a residual set of a plurality of history periods; y2, for each type of period length, constructing a residual error statistical curve based on a historical period residual error set of each type of period length; Y3, the period correction module extracts a period label corresponding to the prediction task according to the timestamp information in the input characteristic of the current moment; y4, based on the time period label, matching and selecting a history period residual template which is most similar to the current prediction input from the residual statistical curve; y5, carrying out correction operation on each time point residual value in the selected period residual template in one-to-one correspondence with the first-round prediction result: corrected predicted value = first round predicted value + corresponding time point residual value; y6, outputting the corrected predicted electric quantity sequence after the period residual errors are aligned as the output of a period correction module; The electricity selling plan generating module divides the whole electricity selling contract performance electric quantity into a plurality of fine-grained fragments, and dynamically judges whether each fragment enters an execution state according to the prediction deviation of each fragment; The method for judging the prediction deviation dynamic state by the electricity selling plan generating module comprises the following steps: a1, performing cycle of target electricity selling contract According to the granularity of the preset time Division into n= / Each long-time segment, get the contract segment set C= { , ,..., And (3) a process for preparing the same, wherein, =( , ) Representing the ith contract segment, having an execution period And target performance power ; A2, obtaining segment predicted values, and extracting and obtaining segment time periods from the final predicted electric quantity sequence output by the disturbance feedback module Corresponding predicted value I.e. Wherein, the method comprises the steps of, Representing prediction model versus time period Is used for predicting the electric quantity value of the power supply; A3, calculating a prediction error for each contract segment by the system: at the same time set a deviation threshold ; A4 for each fragment : If it is ≤ Then the fragment is marked as "executable"; If it is > Setting the fragment state as 'frozen', not entering into execution scheduling, and pushing the fragment state into a queue Q to be deferred; the frozen fragments were recorded as q= { ∣ > }; A5, continuously predicting again in a subsequent time period by the system; recalculating fragments in frozen fragment queue Q every time a new prediction is updated And making a judgment: If it meets ≤ Thawing the fragment, and adding the latest executable period for redemption; If the duration does not meet the deviation requirement and exceeds the maximum delay window, entering a manual scheduling or default processing flow; a6, when settlement is carried out, the actual execution electric quantity is carried out by taking contract fragments as units And the target power consumption Calculating deviation and recording deviation rate : ; And the system is based on the deviation rate of each executed contract segment Respectively executing grading reward and punishment settlement with a preset reward and punishment segmentation strategy; the electric quantity prediction module detects a trend mutation point in an extended input sequence, selects a target prediction model from a plurality of preset candidate prediction models based on the current input characteristic state to perform electric quantity prediction, and comprises the following steps: H1, receiving in real time the extended input sequence output by the input construction module , An extended input vector representing the current time t; setting the length W of a sliding window, and calculating the slope, gradient and residual standard deviation of the key input characteristic dimension in a continuous time step; H2, when the change rate of a certain key input characteristic dimension is monitored to exceed a preset slope threshold value or the residual standard deviation exceeds a residual mutation threshold value, the system judges that the current input sequence has trend bifurcation and generates a bifurcation zone bit ; H3, the system maintains a plurality of candidate prediction models, when the bifurcation is marked When the system carries out matching evaluation on each candidate model through a similarity scoring function according to the context state of the current input feature, and selects a target prediction model which is most similar to the current state from the candidate models to predict; After H4, model reconstruction is completed, a predicted electric quantity sequence is generated Wherein H is the prediction step length; At the same time, the bifurcation is marked And the number of the selected model and the information of the switching time stamp are transmitted to the disturbance feedback module.
2. 2. The electricity management system of claim 1, wherein the H4 specifically comprises: After receiving the model switching control instruction, the system loads the corresponding candidate prediction model from the model library according to the selected model identification ; The system outputs the extended input sequence output by the input construction module As input to a target prediction model; The target prediction model receives the extended input sequence Then, executing a complete forward propagation operation, and recursively outputting an electric quantity predicted value in a future predicted time window through a multi-layer structure of the neural network; the output of the target prediction model is a power prediction sequence in a prediction time range [ t+1, t+H ], and the power prediction sequence is expressed as follows: Wherein, the method comprises the steps of, Representing a predicted value of the electrical quantity for a future h time step.

Description

Electricity selling management system capable of predicting electric quantity Technical Field The invention relates to the technical field of electricity selling management, in particular to an electricity selling management system capable of predicting electric quantity. Background In modern power systems, power prediction is an important basis for power scheduling and market trading. The traditional electric quantity prediction method mainly depends on historical load data and basic environmental factors, short-term or medium-long-term prediction is realized by establishing a statistical regression model or a time sequence model, and the method can play a good role in processing load scenes with strong regularity and obvious periodicity, so that the running stability of an electric power system and the basic requirements of an electric power market are supported; With the gradual increase of renewable energy access proportion and the diversification of load characteristics at the user side, electric quantity prediction faces more uncertainties, such as extreme meteorological events caused by climate change, stepwise changes of user behaviors, dynamic adjustment of policies and electricity price mechanisms, all of which can cause load data to present new complex characteristics in the aspects of trend, fluctuation and sudden disturbance; When the existing method based on single time sequence or traditional regression analysis is used for coping with multidimensional disturbance conditions, the prediction precision and the adaptability of the existing method have certain limitations; In order to improve accuracy and robustness of electric quantity prediction, researchers gradually introduce a deep learning method and a multi-source data fusion idea, and attempt to capture a complex nonlinear relationship by using a neural network model. The method can improve the prediction performance to a certain extent, but still needs to solve two key problems in long-term operation, namely, how to enhance the structural expression capability of input features so as to better identify trend inversion and cycle jump; Therefore, the application further provides an electric quantity prediction system taking the period correction and the disturbance feedback mechanism into consideration, which has stronger practical significance and application value, and further provides an electricity selling management system capable of predicting electric quantity. Disclosure of Invention The present invention is directed to an electricity selling management system capable of predicting electricity quantity, so as to solve the above-mentioned problems in the background art. The invention can realize the electricity selling management system capable of predicting the electric quantity, which comprises a data acquisition module, an input construction module, an electric quantity prediction module, a period correction module, a disturbance feedback module and an electricity selling plan generation module; the data acquisition module is used for acquiring basic data required by prediction, wherein the basic data comprises historical electric quantity data, environment data, running state data, contract information data and scheduling event data; the input construction module is used for preprocessing and structurally enhancing basic data and comprises the following steps: extracting and forming a preliminary input feature sequence from basic data, wherein the preliminary input feature sequence comprises a plurality of feature vectors, and each feature vector represents an input state under a time segment and is called an original input vector; after normalization and time alignment processing are carried out on the original input vector, the input construction module generates a corresponding mirror image vector for preset key input features (such as temperature and load) according to preset symmetrical structure construction rules, wherein the mirror image vector is used for reflecting symmetrical behavior of the input features under the condition of possible trend inversion; the input construction module splices the original input vector and the mirror image vector to form an extended input sequence, so that the recognition capability of the model on the trend mutation points is enhanced; The electric quantity prediction module is used for receiving the extended input sequence, generating a first round of prediction result through the time sequence neural network model, and outputting the first round of prediction electric quantity sequence in a prediction period; The period correction module performs trend adjustment on the first-round predicted electric quantity sequence based on the historical periodic error behavior; The cycle correction module establishes a plurality of residual templates with fixed cycle length, such as 7-day cycle, 14-day cycle and 30-day cycle, establishes a residual statistical curve for each cycle, and matches t