A power load forecasting method and apparatus

Abstract

The application provides a power load prediction method and device, and belongs to the technical field of power load prediction.The method comprises the following steps: acquiring current waveform data, and extracting fundamental wave and harmonic wave components; performing waveform space form geometry analysis to obtain a real-time load characteristic sequence, and then performing segmentation processing; converting current effective value sequences of each time window into time-frequency energy distribution vectors to construct a three-level characteristic library; constructing a three-dimensional tensor model through device start-stop event identification, and then inputting the three-dimensional tensor model into a multi-target optimizer to evolve characteristic weights, filtering abnormal samples to obtain a characteristic cluster; performing random masking processing on time sequence data of the characteristic cluster to generate a mask sequence, inputting the mask sequence into an encoder to reconstruct the masked data, comparing and learning to determine abnormalities, outputting modified data, and then inputting the modified data into a prediction network to generate a feedback signal stream; and analyzing the feedback signal stream to update prediction network weights.Based on the method, a power load prediction device is further provided.The application significantly improves the accuracy of power load prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power load prediction, and particularly relates to a power load prediction method and device. BACKGROUND

[0002] In the efficient operation and fine management process of the smart grid, accurate power load prediction is the core link of optimizing resource scheduling and guaranteeing power supply stability. With the wide deployment of terminal devices such as smart meters, the power grid system can collect massive raw current waveform data at a frequency of seconds, which contains rich load characteristic information, including real-time parameters such as voltage effective value, active power, and reactive power, providing a data basis for high-precision load prediction.

[0003] Traditional load prediction methods, such as time series analysis based on statistics and classic machine learning algorithms, often fail to accurately capture load variation patterns when dealing with high-dimensional, nonlinear, and complex time-varying power consumption data. On the one hand, the detailed characteristics such as harmonic components and device start-stop transient responses in the original current waveform data are not fully explored, resulting in incomplete load feature extraction. On the other hand, the randomness of user power consumption behavior and the dynamic influence of meteorological factors on load make it difficult for traditional models to establish accurate correlations. In recent years, deep learning technology has emerged in the field of load prediction due to its strong feature learning ability, especially recurrent neural networks such as long short-term memory (LSTM), which can effectively process time series data. However, existing prediction models based on LSTM still have limitations. Model parameters usually rely on fixed optimization strategies and cannot be dynamically adjusted according to real-time changes in load characteristics and prediction errors, resulting in a significant decrease in prediction accuracy when facing user power consumption pattern mutations, extreme weather, and other abnormal conditions, making it difficult to meet the real-time and accurate operation requirements of the smart grid. SUMMARY

[0004] To solve the above technical problems, the present application proposes a power load prediction method and device, which significantly improves the accuracy of power load prediction, especially enhances the dynamic adaptability to load fluctuations, and can effectively cope with the time-varying characteristics of power consumption patterns.

[0005] To achieve the above purpose, the present application adopts the following technical solutions:

[0006] A power load prediction method, comprising the following steps:

[0007] Obtain current waveform data of a power system and extract fundamental and harmonic components from the current waveform data; then perform geometric topology analysis of the current waveform space form to obtain real-time load feature sequences;

[0008] The real-time load characteristic sequence is segmented by using a sliding time window, and the current effective value sequence in each time window is converted into a time-frequency energy distribution feature vector; a three-level feature library is constructed based on the frequency energy distribution feature vector, which integrates device fingerprint features, user behavior features and environmental response features;

[0009] Based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lags and power consumption is constructed through device start-stop event recognition, the three-dimensional tensor is input into a multi-objective optimizer to evolve feature weights, spectral clustering is performed based on the evolved optimal weights, and abnormal samples are dynamically filtered to obtain feature clusters with homogeneous power consumption modes;

[0010] Random masking processing is performed on the time series data of the feature clusters to generate a mask sequence, the mask sequence is input into a Transformer encoder to reconstruct the masked data, and abnormality is judged by joint contrast learning, and feature correction data is output;

[0011] The feature correction data is input into a long short-term memory prediction network, a quantile loss function is used to dynamically adjust the confidence interval, and a feedback signal stream is generated when continuous out-of-bounds occurs;

[0012] The feedback signal stream is analyzed, the long short-term memory prediction network weight is updated by using a parameterized noise exploration strategy, and closed-loop optimization of the long short-term memory prediction network is realized.

[0013] The application also provides a power load prediction device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the computer program implements the steps of the method when executed by the processor.

[0014] The effects provided in the summary are only the effects of the embodiments, not all the effects of the application, and one of the above technical solutions has the following advantages or beneficial effects:

[0015] This invention proposes a power load forecasting method and device, belonging to the field of power load forecasting technology. The method includes the following steps: acquiring current waveform data of the power system and extracting the fundamental and harmonic components from the current waveform data; then performing geometric topological analysis of the spatial morphology of the current waveform to obtain a real-time load feature sequence; segmenting the real-time load feature sequence using a sliding time window; and converting the current effective value sequence within each time window into a time-frequency domain energy distribution feature vector; constructing a three-level feature library based on the frequency domain energy distribution feature vector, which integrates equipment fingerprint features, user behavior features, and environmental response features; and based on the three-level feature library, constructing a three-dimensional model of meteorological factors, time lag, and electricity consumption through equipment start-up and shutdown event recognition. A tensor model is used to input the three-dimensional tensor into a multi-objective optimizer to evolve feature weights. Based on the evolved optimal weights, spectral clustering is performed and abnormal samples are dynamically filtered to obtain feature clusters with homogeneous electricity consumption patterns. Random masking processing is applied to the time-series data of the feature clusters to generate a mask sequence. This mask sequence is input into a Transformer encoder to reconstruct the masked data, and contrastive learning is used to identify anomalies, outputting corrected feature data. The corrected feature data is input into a Long Short-Term Memory (LSTM) prediction network, and the confidence interval is dynamically adjusted using a quantile loss function. When the confidence interval is continuously exceeded, a feedback signal stream is generated. The feedback signal stream is analyzed, and a parameterized noise exploration strategy is used to update the LTM prediction network weights, achieving closed-loop optimization of the LTM prediction network. Based on this power load forecasting method, a power load forecasting device is also proposed. This invention achieves refined extraction of load features through second-level high-frequency data acquisition combined with discrete Fourier transform and geometric topology analysis. A three-level feature library is constructed using wavelet packet decomposition and Pearson correlation mapping, providing multi-dimensional basic data support for forecasting. By introducing a quantile loss function to dynamically adjust the confidence level and implementing online closed-loop optimization of the prediction model parameters through a deep deterministic strategy gradient framework, the accuracy of power load forecasting is significantly improved, especially the dynamic adaptability to load fluctuations is enhanced, and it can effectively cope with the time-varying characteristics of electricity consumption patterns.

[0016] This invention leverages dynamic time warping matching and spectral clustering techniques to achieve equipment start-up and shutdown event identification and user homogeneous electricity consumption pattern clustering. Combined with mask sequence reconstruction and contrastive learning enhancement using the Transformer architecture, it significantly improves the ability to identify and correct abnormal electricity consumption samples. This not only helps to uncover patterns in user electricity consumption behavior but also enables the timely detection of equipment failures and other anomalies, providing a reliable basis for the refined management of power systems.

[0017] The application improves the quality and stability of input features by evolutionary optimization of feature weights through the NSGA-II multi-objective optimizer, and combines data repair and abnormality discrimination of the double-check correction module. In the reinforcement learning framework, the prediction error is used as the reward signal, and the adaptive adjustment of the model parameters is realized through the collaborative work of the Actor-Critic network, forming a closed-loop mechanism of "prediction-feedback-optimization", which significantly enhances the robustness and generalization ability of the model in complex power consumption environment.

[0018] The high-precision load prediction result realized by the application can provide scientific decision support for the power dispatch department, help to reasonably plan the power generation plan, optimize the power grid operation mode, and reduce the dispatching pressure brought by the peak-valley difference. At the same time, the in-depth analysis of the user power consumption mode can provide a basis for demand side response strategy making, promote reasonable power consumption of users, improve energy utilization efficiency, and promote the development of the power system in a more economic and more environmentally friendly direction. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A power load prediction method flowchart is provided for the embodiment 1 of the application.

[0020] Figure 2 A power load prediction device schematic diagram is provided for the embodiment 2 of the application. DETAILED DESCRIPTION

[0021] To clearly illustrate the technical features of the present application, the following will describe the application in detail through specific embodiments, and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples to implement the different structures of the application. In order to simplify the disclosure of the application, the components and settings of specific examples are described in the following. In addition, the application can repeatedly refer to the numbers and / or letters in different examples. Such repetition is for the purpose of simplification and clarity, and it does not indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. The application omits the description of known components and processing techniques and processes to avoid unnecessary limitation of the application.

[0022] Embodiment 1

[0023] The embodiment 1 of the application provides a power load prediction method, which is used to solve the technical problem that the load prediction accuracy in the prior art is significantly reduced, and it is difficult to meet the real-time and accuracy operation requirements of the smart grid.

[0024] Figure 1 A power load prediction method flowchart is provided for the embodiment 1 of the application.

[0025] In step S100, the process is started.

[0026] In step S110, current waveform data of the power system is acquired, and fundamental wave components and harmonic wave components in the current waveform data are extracted; then, geometric topology analysis of the current waveform space form is performed to obtain a real-time load characteristic sequence;

[0027] The smart meter terminal collects original current waveform data on the user side at a sampling frequency of seconds;

[0028] The collected original current waveform data is subjected to frequency domain analysis and time domain form analysis in parallel;

[0029] The original current waveform data is processed by discrete Fourier transform (DFT) to accurately separate and extract fundamental wave components and harmonic wave components of the current signal;

[0030] The original current waveform data is subjected to waveform morphological feature extraction, and key waveform parameters are calculated. Waveform morphological feature extraction of the original current waveform data and calculation of key waveform parameters are core links for depicting the instantaneous state of the load from the time domain dimension, and need to combine the periodicity, distortion characteristics of the power waveform and the physical law of device operation to achieve accurate quantization through a multi-step algorithm.

[0031] The pre-processing process is as follows: the original current waveform data (the sampling frequency is seconds, i.e. 1 Hz or higher, depending on the configuration of the smart meter) may contain measurement noise (such as sensor thermal noise) and pulse interference (such as spikes at the moment of switching action), and needs to be pre-processed to ensure the accuracy of feature extraction:

[0032] The noise filtering process is as follows: the original waveform data is subjected to smoothing processing, and the formula is:

[0033] ;

[0034] Wherein, is the original sampling point data; is the filtered data, which is used to suppress high-frequency random noise; for waveforms with obvious spike interference (such as pulses generated by electric welders and relay actions), threshold method is used to eliminate outliers. Set the threshold value to 3 times the standard deviation of the waveform mean, and when a sampling point satisfies ; is the waveform mean; is the standard deviation; replace the point with linear interpolation of the adjacent points.

[0035] For periodic current waveforms (such as sinusoidal fundamental wave components), the waveform period T (usually 50 Hz for city power, period 0.02 s) is determined by zero-crossing point detection, and the continuous waveform is divided into a single period sub-waveform sequence, which facilitates the periodic analysis of the subsequent feature points.

[0036] The core of waveform morphology features is to identify the key feature points (such as wave peak, wave trough, zero-crossing point, inflection point) that can reflect the "shape" of the waveform, providing basic coordinates for parameter calculation.

[0037] The local extremum detection algorithm is used to detect the preprocessed waveform in the sliding window.

[0038] If a point satisfies , it is determined as a wave peak; if it satisfies , it is determined as a wave trough.

[0039] For non-periodic waveforms (such as pulse loads), the gradient change rate needs to be verified: the first derivative (difference) at the wave peak / wave trough should be 0, and the second derivative (second difference) at the wave peak should be negative, and at the wave trough should be positive.

[0040] The zero-crossing point is the time when the current waveform crosses the zero level, reflecting the symmetry of the waveform. The linear interpolation method is used for accurate positioning: when the two consecutive points and (or vice versa), the zero-crossing point time satisfies:

[0041] ;

[0042] The number of zero-crossing points in a single period is counted: pure sine wave (fundamental wave) is 2 (positive half cycle → negative half cycle, negative half cycle → positive half cycle), and additional zero-crossing points may occur for waveforms containing harmonics or distortion (such as a square wave containing 3 harmonics, the number of zero-crossing points increases).

[0043] The inflection point is the point with the maximum change rate of the waveform slope, reflecting the mutation characteristics of the waveform (such as the commutation time of rectifying equipment). Calculate the first difference (slope) of the waveform , and then take the second difference of , the point with the maximum absolute value of the second difference is the inflection point, corresponding to the "steepness mutation point" of the rising edge / falling edge of the waveform.

[0044] Based on the above feature points, the core parameters that can represent the waveform morphology are calculated, covering the amplitude characteristics, distortion characteristics, and change rate of the waveform:

[0045] The wave peak factor reflects the deviation of the waveform peak value from the effective value, which is an important indicator of device load impact.

[0046] ;

[0047] where, is the wave peak factor; is the maximum value of the detected wave peak absolute value, is the current effective value.

[0048] The trough depth is the absolute value of the difference between the negative half trough and the zero level, used to evaluate the asymmetry of the positive and negative half of the waveform, and is expressed as: .

[0049] The waveform distortion rate is expressed as:

[0050] ;

[0051] Wherein, is the actual waveform sampling point, is the sampling point of the ideal sinusoidal wave of the same period (consistent with the fundamental frequency), is the number of sampling points of a single period.

[0052] The zero-crossing offset represents the average time difference between the actual zero-crossing point and the ideal sinusoidal wave zero-crossing point, reflecting the phase distortion of the waveform. It is expressed as:

[0053] ;

[0054] Wherein, is the actual zero-crossing time, is the ideal sinusoidal wave zero-crossing time, is the number of zero-crossing points.

[0055] Integrate the fundamental component, harmonic component information extracted by the above frequency domain analysis and the waveform parameters obtained by the time domain shape analysis, fuse and output the load characteristic sequence containing voltage effective value, current effective value, active power, reactive power, harmonic distortion rate and wave crest factor and other key characteristics in real time; The sequence is used as a real-time data stream representing the instantaneous load state.

[0056] In step S120, the real-time load characteristic sequence is segmented by using a sliding time window; and the current effective value sequence in each time window is converted into a time-frequency energy distribution feature vector; based on the frequency domain energy distribution feature vector, a three-level feature library is constructed, which integrates device fingerprint features, user behavior features and environmental response features;

[0057] The real-time load characteristic sequence is segmented and intercepted by using a fixed length sliding window, and the window length covers the complete start-stop cycle of the typical electrical equipment, and the step is 1 / 10 of the window length;

[0058] The current effective value sequence in each window is executed by 6-layer wavelet packet decomposition, db4 wavelet basis function is selected to generate 64 frequency band sub-signals, the energy proportion of each frequency band is calculated and normalized to form a time-frequency energy distribution feature vector;

[0059] Based on the known device operation template library, the Pearson correlation coefficient matrix of the energy distribution of each frequency band and the device type is calculated, and the strongly correlated frequency bands with a correlation coefficient exceeding 0.85 are selected as device fingerprint features;

[0060] Through periodic analysis of the wavelet packet decomposition coefficients, high-frequency energy fluctuation patterns at the day / week scale are identified, and a user behavior time sequence pattern code is constructed in combination with the active power change rate;

[0061] For the characteristic frequency bands corresponding to temperature-sensitive devices (air conditioners, electric heating), the lag cross-correlation function of the energy change and the outdoor temperature is calculated, and the temperature response coefficient and lag period features are extracted;

[0062] Fusion of device fingerprint features, user behavior features and environmental response features, construction of three-level feature library and output to dynamic time warping matching module.

[0063] In step S130, based on the three-level feature library, a three-dimensional tensor model of weather factors, time lag and power consumption is constructed through device start-stop event identification, the three-dimensional tensor is input into a multi-objective optimizer to evolve feature weights, spectral clustering is performed based on the evolved optimal weights, and abnormal samples are dynamically filtered to obtain feature clusters with homogeneous power consumption patterns.

[0064] From the three-level feature library, the device operation feature sequence is extracted, and the dynamic time warping algorithm is used to calculate the similarity distance between the device operation feature sequence and the preset standard device start-stop template. When the similarity exceeds the threshold, the air conditioner compressor start-stop event point is marked;

[0065] For the identified start-stop event, the lag cross-correlation of the power change curve before and after the event and the outdoor temperature time series data is analyzed to determine the temperature lag response time constant;

[0066] Integrate temperature data, lag response time and event-related power consumption to construct a three-dimensional tensor correlation model of weather factors-time lag-power consumption;

[0067] Input the three-dimensional tensor into the NSGA-II multi-objective optimizer, take the silhouette coefficient and Davies-Bouldin index of feature clustering as the dual-objective fitness function, and perform weight vector evolution on temperature sensitivity, load fluctuation period, and power mutation amplitude;

[0068] Simulated binary crossover operator is used to generate offspring weight vectors, and elitist strategy is used to select the Pareto optimal solution set, and after a preset number of iterations, the Pareto front optimal feature weight combination is output;

[0069] The optimal weight combination is used to weight the user feature vector, input into the spectral clustering engine to construct the weighted covariance matrix, and the sample similarity is calculated based on the weighted Mahalanobis distance;

[0070] The Laplace matrix feature decomposition is performed to determine the cluster center, and the Mahalanobis distance of each sample to the cluster center is calculated in real time;

[0071] A dynamic filtering threshold (3 times the standard deviation from the cluster center) is set, and after filtering out abnormal samples, the characteristic cluster with homogeneous power consumption mode is output to the double verification correction module.

[0072] In step S140, random masking processing is performed on the time series data of the feature cluster to generate a mask sequence, the mask sequence is input into the Transformer encoder to reconstruct the masked data, and the abnormality is judged by joint contrast learning, and the feature correction data is output;

[0073] Random masking processing is performed on the time series data of the input feature cluster, and a mask sequence is generated by randomly selecting 20% of the continuous time segments and setting them to zero;

[0074] The mask sequence is input into the Transformer encoder, the context dependency relationship is learned through the multi-head attention layer, the time sequence information is preserved by using the position encoding, and the data of the masked segment is reconstructed;

[0075] The mean square error between the reconstructed data and the original data is calculated as the reconstruction loss, and when the error is lower than the set threshold, it is determined that the sequence reconstruction is effective;

[0076] Construct a contrast learning sample set: extract normal power consumption segments from the device running log as positive samples, and inject known fault modes (voltage drop / harmonic distortion) to generate negative samples;

[0077] Perform negative sampling in the feature space, mix positive and negative samples at a ratio of 1:3, and extract high-dimensional feature vectors through a deep neural network;

[0078] Use the InfoNCE loss function to optimize the similarity in the feature space, maximize the cosine similarity of the positive samples, and minimize the similarity of the negative samples, forming an abnormal pattern discrimination boundary;

[0079] Set the dynamic safety margin δ=0.5·(σ_user+σ_cluster), where σ_user is the historical fluctuation variance of the user, and σ_cluster is the variance within the cluster; inject Gaussian adversarial disturbance to the negative sample to enhance the robustness of the discrimination boundary.

[0080] Perform anomaly discrimination on the reconstructed data: if the feature vector of the sample is less than the safety margin from the anomaly discrimination boundary, it is marked as suspicious data and replaced with the reconstructed value;

[0081] Output the feature correction data that has been double-verified by sequence reconstruction and anomaly discrimination to the long short-term memory prediction network.

[0082] In step S150, the feature correction data is input into the long short-term memory prediction network, a quantile loss function is used to dynamically adjust the confidence interval, and a feedback signal stream is generated when the continuous boundary is crossed.

[0083] A double-channel long short-term memory prediction network is constructed, wherein the main channel is used for inputting the feature correction data, and the auxiliary channel is used for inputting the temperature sensitivity feature.

[0084] The feature correction data is input into the constructed double-channel long short-term memory prediction network; a quantile loss function is used to dynamically calculate the prediction interval coverage probability, and the confidence level is adaptively adjusted according to the load fluctuation characteristics of the time period; a quantile loss function is used to synchronously output three prediction values: the 10th quantile (lower limit), the 50th quantile (median) and the 90th quantile (upper limit);

[0085] The prediction interval coverage probability (PICP) is dynamically calculated: the proportion of the measured value falling into the [lower limit, upper limit] interval is calculated.

[0086] The confidence level is adaptively adjusted according to the time period characteristics: in the peak period (large load fluctuation), the confidence level is maintained at 90%; in the flat period, the confidence level is reduced to 85%; in the valley period (load stable), the confidence level is reduced to 80%.

[0087] When the measured value exceeds the upper limit of the prediction interval for three consecutive times, the error accumulator is activated to generate a feedback signal stream containing the deviation direction and amplitude. The deviation direction includes upward and downward deviation; the amplitude includes the absolute difference between the measured value and the upper limit; the time period characteristics are coded.

[0088] The cumulative deviation feature is coded as a three-tuple sequence of [direction flag, amplitude value, time period code].

[0089] In step S160, the feedback signal stream is analyzed, the long short-term memory prediction network weight is updated using the parameterized noise exploration strategy, and the long short-term memory prediction network closed-loop optimization is realized.

[0090] The feedback signal stream is input into the deep deterministic policy gradient optimization framework, and the following closed-loop optimization operations are performed.

[0091] The deviation direction and amplitude in the feedback signal stream are analyzed through the action network to generate a long short-term memory (LSTM) prediction network hidden layer weight adjustment strategy;

[0092] The value function of the weight adjustment strategy evaluated by the critic network is calculated to calculate the strategy optimization benefit;

[0093] The prediction error negative logarithm is taken as the immediate reward signal, the experience replay mechanism is used to update the parameters of the Actor network and the Critic network; the sampling priority P(z)=|TD error|+ξ is used.

[0094] Real-time dynamic updating of the LSTM prediction model parameters is realized, and adaptive closed-loop feedback optimization of the prediction error to the model parameters is completed.

[0095] In step S170, the flow ends.

[0096] The power load prediction method proposed in embodiment 1 of the present application realizes fine extraction of load characteristics through high-frequency data acquisition at the second level combined with discrete Fourier transform and geometric topology analysis; three-level feature library is constructed by using wavelet packet decomposition and Pearson correlation mapping, providing multi-dimensional basic data support for prediction. The quantile loss function is introduced to dynamically adjust the confidence level, and the online closed-loop optimization of the prediction model parameters is realized through the deep deterministic policy gradient framework, which significantly improves the accuracy of power load prediction, especially enhances the dynamic adaptability to load fluctuations, and can effectively cope with the time-varying characteristics of power consumption mode.

[0097] The power load prediction method proposed in embodiment 1 of the present application realizes device start-stop event recognition and user homogeneous power consumption mode clustering by means of dynamic time warping matching and spectral clustering technology, and greatly improves the recognition and correction ability of abnormal power consumption samples by combining mask sequence reconstruction and contrast learning enhancement of the Transformer architecture. This not only helps to mine user power consumption behavior rules, but also can timely discover abnormal conditions such as device failure, providing a reliable basis for fine management of the power system.

[0098] The power load prediction method proposed in embodiment 1 of the present application realizes evolution optimization of feature weights by NSGA-II multi-objective optimizer, and improves the quality and stability of input features through data repair and abnormality discrimination of the double-check correction module. In the reinforcement learning framework, the prediction error is used as the reward signal, and the adaptive adjustment of the model parameters is realized through the cooperative work of the Actor-Critic network, forming a closed-loop mechanism of "prediction-feedback-optimization", which significantly enhances the robustness and generalization ability of the model in complex power consumption environment.

[0099] The power load prediction method proposed in embodiment 1 of the present application realizes high-precision load prediction results, which can provide scientific decision support for power dispatching departments, help to reasonably plan power generation, optimize power grid operation mode, and reduce the dispatching pressure brought by peak-valley difference. At the same time, in-depth analysis of user power consumption mode can provide basis for demand side response strategy formulation, promote reasonable power consumption of users, improve energy utilization efficiency, and promote the development of power system to be more economical and more environmentally friendly.

[0100] Embodiment 2

[0101] The present application also proposes a device, Figure 2The power load prediction device for the embodiment 2 of the present application comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the computer program realizes the method steps shown in the figure when executed by the processor Figure 1

[0102] In step S100, the process is started.

[0103] In step S110, current waveform data of the power system is acquired, and fundamental wave components and harmonic components in the current waveform data are extracted; then geometric topology analysis of the current waveform space form is performed to obtain a real-time load characteristic sequence.

[0104] In step S120, the real-time load characteristic sequence is segmented by using a sliding time window; the current effective value sequence in each time window is converted into a time-frequency energy distribution feature vector; and a three-level feature library fusing device fingerprint features, user behavior features and environmental response features is constructed based on the frequency energy distribution feature vector.

[0105] In step S130, based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lags and power consumptions is constructed by device start-stop event recognition, the three-dimensional tensor is input into a multi-objective optimizer to evolve feature weights, spectral clustering is performed based on the evolved optimal weights, and abnormal samples are dynamically filtered to obtain a feature cluster with a homogeneous power consumption mode.

[0106] In step S140, random masking processing is performed on the time sequence data of the feature cluster to generate a mask sequence, the mask sequence is input into a Transformer encoder to reconstruct the masked data, and abnormality is judged by joint contrast learning, and feature correction data is output.

[0107] In step S150, the feature correction data is input into a long short-term memory prediction network, a quantile loss function is used to dynamically adjust the confidence interval, and a feedback signal stream is generated when continuous out-of-bounds occurs.

[0108] In step S160, the feedback signal stream is analyzed, a parameterized noise exploration strategy is used to update the long short-term memory prediction network weight, and closed-loop optimization of the long short-term memory prediction network is realized.

[0109] In step S170, the process is ended.

[0110] ​The power load prediction device provided in Embodiment 2 of the present application realizes fine extraction of load characteristics by high-frequency data acquisition at a second level combined with discrete Fourier transform and geometric topology analysis; three-level feature library is constructed by using wavelet packet decomposition and Pearson correlation mapping, thereby providing multi-dimensional basic data support for prediction. A quantile loss function is introduced to dynamically adjust the confidence level, and online closed-loop optimization of the prediction model parameters is realized through a deep deterministic policy gradient framework, thereby significantly improving the accuracy of power load prediction, especially enhancing the dynamic adaptability to load fluctuations, and effectively coping with the time-varying characteristics of power consumption modes.

[0111] The power load prediction device provided in Embodiment 2 of the present application realizes device start-stop event recognition and user homogeneous power consumption mode clustering by means of dynamic time warping matching and spectral clustering technology, and greatly improves the recognition and correction ability of abnormal power consumption samples by combining mask sequence reconstruction and contrast learning enhancement of the Transformer architecture. This not only helps to mine user power consumption behavior rules, but also can timely discover abnormal conditions such as device failure, thereby providing a reliable basis for fine management of the power system.

[0112] The power load prediction device provided in Embodiment 2 of the present application realizes device start-stop event recognition and user homogeneous power consumption mode clustering by means of dynamic time warping matching and spectral clustering technology, and greatly improves the recognition and correction ability of abnormal power consumption samples by combining mask sequence reconstruction and contrast learning enhancement of the Transformer architecture. This not only helps to mine user power consumption behavior rules, but also can timely discover abnormal conditions such as device failure, thereby providing a reliable basis for fine management of the power system.

[0113] The power load prediction device provided in Embodiment 2 of the present application realizes device start-stop event recognition and user homogeneous power consumption mode clustering by means of dynamic time warping matching and spectral clustering technology, and greatly improves the recognition and correction ability of abnormal power consumption samples by combining mask sequence reconstruction and contrast learning enhancement of the Transformer architecture. This not only helps to mine user power consumption behavior rules, but also can timely discover abnormal conditions such as device failure, thereby providing a reliable basis for fine management of the power system.

[0114] It should be noted that the technical scheme of the present application also provides an electronic device, which comprises: a communication interface capable of interacting with other devices such as network devices; a processor connected with the communication interface to realize information interaction with other devices, used to run a computer program to execute the power load prediction method provided by one or more technical schemes, and the computer program is stored on a memory. Of course, in actual application, various components in the electronic device are coupled together through a bus system. It can be understood that the bus system is used to realize the connection and communication between the components. In addition to the data bus, the bus system also includes a power bus, a control bus and a state signal bus. The memory in the embodiment of the present application is used to store various types of data to support the operation of the electronic device. Examples of these data include any computer programs used to operate on the electronic device. It can be understood that the memory can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM, Read Only Memory), a programmable read-only memory (PROM, Programmable Read-Only Memory), an erasable programmable read-only memory (EPROM, Erasable Programmable Read-Only Memory), an electrically erasable programmable read-only memory (EEPROM, Electrically Erasable Programmable Read-Only Memory), a ferromagnetic random access memory (FRAM, ferromagnetic random access memory), a flash memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM, Compact Disc Read-Only Memory); the magnetic surface memory can be a disk memory or a tape memory. The volatile memory can be a random access memory (RAM, Random Access Memory) used as an external cache.By way of example, and not limitation, many forms of RAM can be used, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDR SDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), Direct Rambus Random Access Memory (DRRAM). The memory described in the embodiments of the present application is intended to include, but not be limited to, these and any other suitable types of memory. The methods disclosed in the embodiments of the present application can be applied in or implemented by a processor. The processor can be an integrated circuit chip chip with a processing capability. In the implementation process, each step of the above method can be completed by integrated logic circuits or instructions in the form of software in the processor. The processor described above can be a general-purpose processor, a DSP (Digital Signal Processing, i.e., a chip capable of implementing digital signal processing technology), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The processor can implement or execute the methods, steps and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor can be a microprocessor or any conventional processor. In combination with the steps of the method disclosed in the embodiments of the present application, the hardware decoding processor can be directly embodied to execute the completion, or the hardware and software modules in the decoding processor are combined to execute the completion. The software module can be located in a storage medium, which is located in the memory, and the processor reads the program in the memory to complete the steps of the above method in combination with the hardware. The processor executes the program to implement the corresponding flow in each method of the embodiments of the present application, and for the sake of brevity, it will not be repeated here.

[0115] The related part of the power load prediction device provided by the embodiment 2 of the present application can refer to the detailed description of the corresponding part of the power load prediction method provided by the embodiment 1 of the present application, and will not be repeated here.

[0116] It should be noted that, in this paper, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device inherent in the series of elements. Without more limitations, the element defined by the sentence "includes a" does not exclude the existence of other same elements in the process, method, article or device including the element. In addition, the above technical solutions provided by the embodiments of the present application have not been described in detail, so as not to be too repetitive.

[0117] The above describes the specific embodiments of the present application in combination with the drawings, but is not a limitation on the protection scope of the present application. For those skilled in the art, on the basis of the above description, other different forms of modification or deformation can also be made. Here, it is not necessary and impossible to enumerate all the embodiments. Various modifications or deformations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A power load forecasting method characterized by, The method comprises the following steps: Obtain current waveform data of a power system, and extract fundamental wave components and harmonic wave components in the current waveform data; then perform geometric topology analysis on the current waveform space form to obtain a real-time load characteristic sequence; Segment the real-time load characteristic sequence using a sliding time window; and convert the current effective value sequence in each time window into a time-frequency energy distribution feature vector; construct a three-level feature library that fuses device fingerprint features, user behavior features, and environmental response features based on the frequency energy distribution feature vector; Based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lags, and power consumption is constructed through device start-stop event identification; the three-dimensional tensor is input into a multi-objective optimizer to evolve feature weights; based on the evolved optimal weights, spectral clustering is performed and abnormal samples are dynamically filtered to obtain feature clusters with homogeneous power consumption patterns; Perform random masking processing on the time series data of the feature clusters to generate a mask sequence; input the mask sequence into a Transformer encoder to reconstruct the masked data; and jointly perform contrastive learning to identify abnormalities; output feature correction data; Input the feature correction data into a long short-term memory prediction network; dynamically adjust the confidence interval using a quantile loss function; and generate a feedback signal stream when consecutive out-of-bounds occur; Analyze the feedback signal stream; update the long short-term memory prediction network weights using a parameterized noise exploration strategy; and realize closed-loop optimization of the long short-term memory prediction network.

2. The power load forecasting method of claim 1, wherein, The real-time load characteristic sequence is used to represent real-time data streams of instantaneous load states, including voltage effective value, current effective value, active power, reactive power, harmonic distortion rate, and wave crest factor.

3. The method of claim 1, wherein, Segment the real-time load characteristic sequence using a sliding time window; and convert the current effective value sequence in each time window into a time-frequency energy distribution feature vector; specifically: Segment and intercept the real-time load characteristic sequence using a fixed-length sliding window; and the window length covers the complete start-stop cycle of the power consumption device; Perform 6-layer wavelet packet decomposition on the current effective value sequence in each time window; select db4 wavelet basis function to generate 64 frequency band sub-signals; calculate the energy proportion of each frequency band and normalize it to form a time-frequency energy distribution feature vector.

4. The method of claim 1, wherein, Based on the frequency energy distribution feature vector, construct a three-level feature library that fuses device fingerprint features, user behavior features, and environmental response features; specifically: Calculate the Pearson correlation coefficient matrix of the energy distribution of each frequency band and the device type; select the correlation frequency bands with a Pearson correlation coefficient greater than a threshold value as the device fingerprint features; Identify the high-frequency energy fluctuation pattern at the preset time scale in the frequency energy distribution feature vector; and construct the user behavior features in combination with the active power change rate; Calculate the lag cross-correlation function between the energy change of the characteristic frequency band of the temperature-sensitive device and the outdoor temperature to obtain the environmental response features.

5. The method of claim 1, wherein, Based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lags, and power consumption is constructed through device start-stop event identification; the three-dimensional tensor is input into a multi-objective optimizer to evolve feature weights; based on the evolved optimal weights, spectral clustering is performed and abnormal samples are dynamically filtered to obtain feature clusters with homogeneous power consumption patterns; Specifically: The device start-stop event includes an air conditioner compressor start-stop event; the environmental factor lag response time is a temperature lag response time constant; the spectral clustering engine determines a cluster center by eigen decomposition of a Laplacian matrix, and an abnormal sample is a sample outside 3 times the standard deviation from the cluster center. The time series data of the feature cluster is subjected to random masking processing to generate a mask sequence, which is input into a Transformer encoder to reconstruct the masked data, and joint contrast learning is used to determine the anomaly, and the feature correction data is output, specifically: Randomly mask the time series data of the input feature cluster, and randomly select a continuous time segment according to a preset proportion to generate a mask sequence; 6. The method of claim 5, wherein, Input the mask sequence into the Transformer encoder to learn the context dependency through the multi-head attention layer, and use the position encoding to retain the time series information to reconstruct the data of the masked segment; 7. The method of claim 1, wherein, When the reconstructed data of the masked segment is effective, the normal power consumption segment and the fault power consumption segment are negatively sampled in the feature space, the InfoNCE loss function is used to optimize the abnormal mode discrimination boundary, and the feature correction data that has been verified by sequence reconstruction and anomaly discrimination is output. The feature correction data is input into a long short-term memory prediction network, and a quantile loss function is used to dynamically adjust the confidence interval, and a feedback signal stream is generated when consecutive out-of-bounds occurs, specifically: A double-channel long short-term memory prediction network is constructed, in which the main channel is used to input the feature correction data, and the auxiliary channel is used to input the temperature sensitivity feature; The feature correction data is input into the constructed double-channel long short-term memory prediction network; a quantile loss function is used to dynamically calculate the prediction interval coverage probability, and the confidence level is adaptively adjusted according to the time-of-use load fluctuation characteristics; 8. The method of claim 1, wherein, When the measured value exceeds the upper limit of the prediction interval for a preset number of consecutive times, the error accumulator is activated to generate a feedback signal stream containing the deviation direction and amplitude. The feedback signal stream is analyzed, and the long short-term memory prediction network weight is updated using a parameterized noise exploration strategy to realize closed-loop optimization of the long short-term memory prediction network, specifically: The deviation direction and amplitude in the feedback signal stream are analyzed through the actor network to generate a long short-term memory prediction network hidden layer weight adjustment strategy The critic network evaluates the strategy value function to take the negative logarithm of the prediction error as the immediate reward; 9. The method of claim 1, wherein, The experience replay mechanism is used to update the long short-term memory prediction network parameters to realize online closed-loop adaptive optimization of the long short-term memory prediction network parameters. including: ​ ​ 10. A power load forecasting device characterized by comprising: ​ A memory, a processor, and a computer program stored on the memory and runnable on the processor, which, when executed by the processor, implement the steps of the method of any one of claims 1 to 9.

Images