Power load prediction method and device

Abstract

The invention provides a power load prediction method and device, and belongs to the technical field of power load prediction.The method comprises the steps that current waveform data are obtained, and fundamental wave and harmonic components in the current waveform data are extracted; carrying out waveform spatial form geometric analysis to obtain a real-time load characteristic sequence, and then carrying out segmentation processing; the current effective value sequence of each time window is converted into a time-frequency domain energy distribution vector, and then a three-level feature library is constructed; constructing a three-dimensional tensor model through equipment start-stop event identification, inputting the three-dimensional tensor model into a multi-target optimizer to evolve feature weights, and filtering abnormal samples to obtain a feature cluster; performing random masking processing on the time sequence data of the feature cluster to generate a mask sequence, inputting the mask sequence into an encoder to reconstruct masking data, comparing, learning and judging abnormal output correction data, and inputting the corrected data into a prediction network to generate a feedback signal flow; and analyzing the feedback signal flow to update the prediction network weight. Based on the method, the invention also provides power load prediction equipment. According to the invention, the precision of power load prediction is obviously improved.

Description

Technical Field

[0001] This invention belongs to the field of power load forecasting technology, and specifically relates to a power load forecasting method and device. Background Technology

[0002] In the process of efficient operation and refined management of smart grids, accurate power load forecasting is a core element in optimizing resource scheduling and ensuring power supply stability. With the widespread deployment of terminal devices such as smart meters, the power grid system can collect massive amounts of raw current waveform data at a frequency of seconds. This data contains rich load characteristic information, including real-time parameters such as effective voltage, active power, and reactive power, providing a data foundation for high-precision load forecasting.

[0003] Traditional load forecasting methods, such as statistical time series analysis and classical machine learning algorithms, often struggle to accurately capture load variation patterns when dealing with high-dimensional, nonlinear, and complex time-varying electricity consumption data. On one hand, detailed features such as harmonic components and transient responses during equipment start-up and shutdown in the raw current waveform data are not fully extracted, resulting in incomplete load feature extraction. On the other hand, the randomness of user electricity consumption behavior and the dynamic impact of meteorological factors on the load make it difficult for traditional models to construct accurate correlations. In recent years, deep learning technology has emerged in the field of load forecasting due to its powerful feature learning capabilities, especially recurrent neural networks such as Long Short-Term Memory (LSTM), which can effectively process time-series data. However, existing LSTM-based forecasting models still have limitations. Model parameters typically rely on fixed optimization strategies and cannot be dynamically adjusted according to real-time changes in load characteristics and prediction errors. This leads to a significant decrease in prediction accuracy when facing abnormal operating conditions such as sudden changes in user electricity consumption patterns and extreme weather, making it difficult to meet the real-time and accuracy requirements of smart grid operations. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a power load forecasting method and device, which significantly improves the accuracy of power load forecasting, and in particular enhances the dynamic adaptability to load fluctuations, effectively addressing the time-varying characteristics of electricity consumption patterns.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting electricity load includes the following steps: Acquire current waveform data of the power system and extract the fundamental and harmonic components from the current waveform data; then perform geometric topological analysis of the spatial morphology of the current waveform to obtain the real-time load characteristic sequence. The real-time load characteristic sequence is segmented using a sliding time window; and the effective current value sequence within each time window is converted into a time-frequency domain energy distribution feature vector; based on the frequency domain energy distribution feature vector, a three-level feature library is constructed that integrates device fingerprint features, user behavior features, and environmental response features; Based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lag, and electricity consumption is constructed by identifying equipment start-up and shutdown events. 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 electricity consumption patterns. Random masking is performed on the temporal data of feature clusters to generate a mask sequence. The mask sequence is then input into the Transformer encoder to reconstruct the masked data. Anomalies are identified through joint contrastive learning, and feature correction data is output. The feature-corrected data is input into a long short-term memory 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 flow is analyzed, and a parameterized noise exploration strategy is used to update the weights of the long short-term memory prediction network, thereby achieving closed-loop optimization of the long short-term memory prediction network.

[0006] The present invention also proposes an electrical load forecasting device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method described.

[0007] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects: 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.

[0008] 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.

[0009] This invention improves the quality and stability of input features by using the NSGA-II multi-objective optimizer to perform evolutionary optimization of feature weights, combined with data repair and anomaly detection by a dual-verification correction module. Within the reinforcement learning framework, prediction error is used as the reward signal, and the collaborative work of the Actor-Critic network enables adaptive adjustment of model parameters, forming a closed-loop mechanism of "prediction-feedback-optimization," which significantly enhances the model's robustness and generalization ability in complex power environments.

[0010] The high-precision load forecasting results achieved by this invention can provide scientific decision support for power dispatching departments, helping to rationally plan power generation schedules, optimize grid operation, and reduce dispatching pressure caused by peak-valley differences. Simultaneously, in-depth analysis of user electricity consumption patterns can provide a basis for demand-side response strategy formulation, promoting rational electricity use by users, improving energy efficiency, and driving the power system towards a more economical and environmentally friendly direction. Attached Figure Description

[0011] Figure 1 This is a flowchart of a power load forecasting method proposed in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of a power load prediction device proposed in Embodiment 2 of the present invention. Detailed Implementation

[0012] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0013] Example 1 Embodiment 1 of this invention proposes a power load forecasting method to solve the technical problem that the power load forecasting accuracy in the prior art is significantly reduced, making it difficult to meet the real-time and accuracy operation requirements of smart grids.

[0014] Figure 1 This is a flowchart of a power load forecasting method proposed in Embodiment 1 of the present invention; In step S100, the process begins.

[0015] In step S110, the current waveform data of the power system is acquired, and the fundamental and harmonic components in the current waveform data are extracted; then, the geometric topology analysis of the spatial morphology of the current waveform is performed to obtain the real-time load characteristic sequence. The smart meter terminal collects raw current waveform data from the user side at a sampling frequency of seconds. Frequency domain analysis and time domain morphological analysis are performed in parallel on the acquired raw current waveform data; The original current waveform data is processed by Discrete Fourier Transform (DFT) to accurately separate and extract the fundamental component and harmonic components of the current signal. The raw current waveform data undergoes waveform morphological feature extraction, and key waveform parameters are calculated. Extracting waveform morphological features and calculating key waveform parameters from the raw current waveform data is a core step in characterizing the instantaneous state of the load from the time domain perspective. This requires combining the periodicity and distortion characteristics of the power waveform with the physical laws governing equipment operation, and employing a multi-step algorithm to achieve accurate quantification.

[0016] The preprocessing process is as follows: The raw current waveform data (sampling frequency on the order of seconds, i.e., 1Hz or higher, depending on the smart meter configuration) may contain measurement noise (such as sensor thermal noise) and impulse interference (such as spikes at the moment of switching action), which need to be preprocessed to ensure the accuracy of feature extraction. The noise filtering process involves smoothing the original waveform data using the following formula: ; in, This is the original sampling point data; The filtered data is used to suppress high-frequency random noise. For waveforms with obvious spike interference (such as pulses generated by welding machines or relay operations), outliers are removed using a threshold method. The threshold is set to three times the standard deviation of the waveform mean. When a certain sampling point... satisfy ; The mean value of the waveform; The standard deviation is used; the point is replaced with the linear interpolation of the adjacent points.

[0017] For periodic current waveforms (such as the fundamental component of a sine wave), the waveform period T (typically 50Hz for mains power, with a period of 0.02s) is determined by zero-crossing detection. This divides the continuous waveform into a sequence of sub-waveforms with a single period, facilitating subsequent periodic analysis of feature points.

[0018] The core of waveform morphology features is to identify key feature points that reflect the "shape" of the waveform (such as peaks, troughs, zero crossings, and inflection points), providing basic coordinates for parameter calculation.

[0019] The preprocessed waveform is processed using a local extremum detection algorithm. Inside the sliding window.

[0020] If a certain point satisfy Then it is determined to be a peak; if it satisfies If it is, then it is determined to be a trough.

[0021] For non-periodic waveforms (such as pulse loads), it is necessary to verify the gradient rate of change: the first derivative (difference) at the peak / trough should be 0, and the second derivative (difference) should be negative at the peak and positive at the trough.

[0022] The zero-crossing point is the moment when the current waveform crosses the zero level, reflecting the symmetry of the waveform. Precise positioning is achieved using linear interpolation: when two consecutive points... And and (Or conversely) at the time of midnight satisfy: ; The number of zero crossings within a single cycle is counted: for a pure sine wave (fundamental wave), there are 2 (positive half-cycle → negative half-cycle, negative half-cycle → positive half-cycle). Waveforms containing harmonics or distortion may have additional zero crossings (e.g., when a square wave contains the 3rd harmonic, the number of zero crossings increases).

[0023] The inflection point is the point where the rate of change of the waveform's slope is greatest, reflecting the abrupt change characteristics of the waveform (such as the commutation moment of a rectifier). The first-order difference (slope) of the waveform is calculated. And then Find the second difference. The point with the largest absolute value of the second difference is the inflection point, which corresponds to the "sudden change in steepness" of the waveform's rising / falling edge.

[0024] Based on the above feature points, the core parameters that characterize the waveform morphology are calculated, covering dimensions such as waveform amplitude characteristics, distortion characteristics, and rate of change: The crest factor reflects the degree of deviation between the peak value and the effective value of the waveform, and is an important indicator of the load impact of equipment.

[0025] ; in, The peak factor; The maximum value of the absolute value of the detected peak. This is the effective value of the current.

[0026] The trough depth is the absolute value of the difference between the trough of the negative half-cycle and the zero level. It is used to evaluate the asymmetry of the positive and negative half-cycles of the waveform. The trough depth is expressed as: .

[0027] Waveform distortion rate, expressed as: ; in, These are the actual waveform sampling points. These are sampling points for an ideal sine wave with the same period (consistent with the fundamental frequency). This represents the number of sampling points in a single period.

[0028] The zero-crossing offset represents the average time difference between the actual zero-crossing point and the ideal zero-crossing point of the sine wave, reflecting the phase distortion of the waveform. It is expressed as: ; in, For the actual time of midnight, For the zero-crossing point of an ideal sine wave, The number of points crossing zero.

[0029] By combining the fundamental and harmonic component information extracted from the frequency domain analysis and the waveform parameters obtained from the time domain morphology analysis, a load characteristic sequence containing key features such as effective voltage value, effective current value, active power, reactive power, harmonic distortion rate, and crest factor is calculated and output in real time; this sequence serves as a real-time data stream characterizing the instantaneous load state.

[0030] In step S120, the real-time load characteristic sequence is segmented using a sliding time window; and the effective current value sequence within each time window is converted into a time-frequency domain energy distribution feature vector; based on the frequency domain energy distribution feature vector, a three-level feature library is constructed that integrates device fingerprint features, user behavior features, and environmental response features; A fixed-duration sliding window is used to segment and extract the real-time load characteristic sequence. The window length covers the complete start-stop cycle of typical electrical equipment, and the step size is 1 / 10 of the window length. A 6-level wavelet packet decomposition is performed on the current effective value sequence within each window. The 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 domain energy distribution feature vector. Based on the known equipment operation template library, the Pearson correlation coefficient matrix between the energy distribution of each frequency band and the equipment type is calculated, and the strongly correlated frequency bands with a correlation coefficient greater than 0.85 are selected as equipment fingerprint features; By analyzing the periodicity of wavelet packet decomposition coefficients, high-frequency energy fluctuation patterns on daily / weekly scales are identified, and user behavior time-series pattern coding is constructed by combining the active power change rate. For temperature-sensitive equipment (air conditioners, electric heating), calculate the hysteresis cross-correlation function between energy change and outdoor temperature in the corresponding characteristic frequency band, and extract the temperature response coefficient and hysteresis characteristics. By integrating device fingerprint features, user behavior features, and environmental response features, a three-level feature library is constructed and output to the dynamic time warping matching module.

[0031] In step S130, based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lag, and electricity consumption is constructed by identifying equipment start-up and shutdown events. 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 electricity consumption patterns.

[0032] The equipment operation feature sequence is extracted from the three-level feature library, and the similarity distance between it and the preset standard equipment start-up and stop template is calculated using the dynamic time warping algorithm. When the similarity exceeds the threshold, the start-up and stop event point of the air conditioner compressor is marked. For the identified start-stop events, analyze the hysteresis cross-correlation between the power change curves before and after the event and the outdoor temperature time series data to determine the temperature hysteresis response time constant; By integrating temperature data, lag response time, and event-related electricity consumption, a three-dimensional tensor correlation model of meteorological factors, time lag, and electricity consumption is constructed. The three-dimensional tensor is input into the NSGA-II multi-objective optimizer, and the silhouette coefficient and Davidson-Bolding index of feature clustering are used as bi-objective fitness functions to perform weighted vector evolution on three types of features: temperature sensitivity, load fluctuation period, and power mutation amplitude. A simulated binary crossover operator is used to generate offspring weight vectors. Pareto optimal solution sets are selected through an elite retention strategy. After a preset number of iterations, the Pareto front optimal feature weight combination is output. The user feature vector is weighted using the optimal weight combination and input into the spectral clustering engine to construct a weighted covariance matrix. Sample similarity is then calculated based on the weighted Mahalanobis distance. Perform Laplacian matrix eigenvalue decomposition to determine cluster centers and calculate the Mahalanobis distance from each sample to the cluster center in real time; Set a dynamic filtering threshold (3 standard deviations from the cluster center), filter out abnormal samples, and output the feature clusters with homogeneous power consumption patterns to the dual verification correction module.

[0033] In step S140, random masking processing is performed on the temporal 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 anomalies are identified by joint contrastive learning, and feature correction data is output. Random masking is performed on the time series data of the input feature clusters, and continuous time segments are randomly selected at a ratio of 20% and set to zero to generate a mask sequence; The mask sequence is input into the Transformer encoder, and the contextual dependencies are learned through a multi-head attention layer. Positional encoding is used to preserve temporal information and reconstruct the data of the masked segment. The mean squared error between the reconstructed data and the original data is calculated as the reconstruction loss. When the error is lower than a set threshold, the sequence reconstruction is considered valid. Construct a comparative learning sample set: extract normal power consumption segments from the equipment operation log as positive samples, and inject known fault modes (voltage drop / harmonic distortion) to generate negative samples; Negative sampling is performed in the feature space, and positive and negative samples are mixed in a 1:3 ratio. High-dimensional feature vectors are then extracted using a deep neural network. The InfoNCE loss function is used to optimize the feature space similarity, maximizing the cosine similarity of positive sample pairs and minimizing the similarity of negative sample pairs, thus forming an abnormal pattern discrimination boundary. The dynamic safety margin is set to δ=0.5·(σ_user+σ_cluster), where σ_user is the historical variance of users and σ_cluster is the variance within the cluster; Gaussian adversarial perturbations are injected into negative samples to enhance the robustness of the discrimination boundary.

[0034] Perform anomaly detection on the reconstructed data: if the distance between the sample feature vector and the anomaly detection boundary is less than the safety margin, it is marked as suspicious data and replaced with the reconstructed value; The output, modified feature data that has undergone dual verification of sequence reconstruction and anomaly detection is fed into the Long Short-Term Memory prediction network.

[0035] In step S150, the feature correction data is input into the long short-term memory prediction network, and the confidence interval is dynamically adjusted using the quantile loss function. When the confidence interval is continuously exceeded, a feedback signal stream is generated.

[0036] A dual-channel long short-term memory prediction network is constructed, in which the main channel is used to input feature correction data and the auxiliary channel is used to input temperature sensitivity features. The feature-corrected data is input into the constructed dual-channel long short-term memory prediction network; the quantile loss function is used to dynamically calculate the prediction interval coverage probability, and the confidence level is adaptively adjusted according to the time-period load fluctuation characteristics; the quantile loss function is used to simultaneously output three prediction values: the 10th quantile (lower bound), the 50th quantile (median), and the 90th quantile (upper bound). Dynamic calculation of predicted interval coverage probability (PICP): the proportion of measured values ​​falling into the [lower bound, upper bound] interval; The confidence level is adaptively adjusted according to the characteristics of different time periods: during peak hours (large load fluctuations): maintain a 90% confidence level; during off-peak hours: reduce to an 85% confidence level; during off-peak hours (stable load): reduce to an 80% confidence level. When three consecutive measured values ​​exceed the upper limit of the prediction interval, the error accumulator is activated to generate a feedback signal stream containing the direction and magnitude of the deviation. The deviation direction includes upward and downward deviation; the magnitude includes the absolute difference between the measured value and the upper limit; and the time period of occurrence is coded.

[0037] The cumulative deviation feature is encoded as a triplet sequence of [direction indicator, amplitude value, time period code].

[0038] In step S160, the feedback signal stream is parsed, and the weights of the long short-term memory prediction network are updated using a parameterized noise exploration strategy to achieve closed-loop optimization of the long short-term memory prediction network.

[0039] Input the feedback signal stream into the deep deterministic policy gradient optimization framework and perform the following closed-loop optimization operation.

[0040] By analyzing the direction and magnitude of the deviation in the feedback signal stream through the action network, a weight adjustment strategy for the hidden layer of the Long Short-Term Memory (LSTM) prediction network is generated. The optimized return of the strategy is calculated by evaluating the value function of the weighted adjustment strategy through a network of critics. The negative logarithm of the prediction error is used as the immediate reward signal, and the parameters of the Actor network and Critic network are updated using an empirical replay mechanism; the sampling priority is P(z) = |TD error| + ξ. Real-time dynamic updating of LSTM prediction model parameters is achieved, enabling adaptive closed-loop feedback optimization from prediction error to model parameters.

[0041] In step S170, the process ends.

[0042] The power load forecasting method proposed in Embodiment 1 of this invention achieves refined extraction of load characteristics through high-frequency data acquisition at the second level 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. A quantile loss function is introduced to dynamically adjust the confidence level, and an online closed-loop optimization of the forecasting model parameters is achieved through a deep deterministic strategy gradient framework. This significantly improves the accuracy of power load forecasting, especially enhancing the dynamic adaptability to load fluctuations and effectively addressing the time-varying characteristics of electricity consumption patterns.

[0043] The power load forecasting method proposed in Embodiment 1 of this invention utilizes dynamic time warping matching and spectral clustering techniques to achieve equipment start-up and shutdown event identification and user homogeneous power 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 power consumption samples. This not only helps to uncover patterns in user power consumption behavior but also enables the timely detection of abnormal situations such as equipment failures, providing a reliable basis for the refined management of the power system.

[0044] The power load forecasting method proposed in Embodiment 1 of this invention improves the quality and stability of input features by using the NSGA-II multi-objective optimizer to perform evolutionary optimization of feature weights and combining data repair and anomaly detection with a dual-verification correction module. In the reinforcement learning framework, prediction error is used as the reward signal, and the collaborative work of the Actor-Critic network enables adaptive adjustment of model parameters, forming a closed-loop mechanism of "prediction-feedback-optimization," which significantly enhances the robustness and generalization ability of the model in complex power consumption environments.

[0045] The power load forecasting method proposed in Embodiment 1 of this invention provides high-precision load forecasting results, which can offer scientific decision support to power dispatching departments. This helps to rationally plan power generation schedules, optimize grid operation, and reduce dispatching pressure caused by peak-valley differences. Simultaneously, in-depth analysis of user electricity consumption patterns can provide a basis for demand-side response strategy formulation, promote rational electricity use by users, improve energy efficiency, and drive the power system towards a more economical and environmentally friendly direction.

[0046] Example 2 The present invention also proposes a device, Figure 2 This is a schematic diagram of a power load forecasting device according to Embodiment 2 of the present invention, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is executed by the processor to implement... Figure 1 The method steps shown are as follows: In step S100, the process begins.

[0047] In step S110, the current waveform data of the power system is acquired, and the fundamental and harmonic components in the current waveform data are extracted; then, the geometric topology analysis of the spatial morphology of the current waveform is performed to obtain the real-time load characteristic sequence. In step S120, the real-time load characteristic sequence is segmented using a sliding time window; and the effective current value sequence within each time window is converted into a time-frequency domain energy distribution feature vector; based on the frequency domain energy distribution feature vector, a three-level feature library is constructed that integrates device fingerprint features, user behavior features, and environmental response features; In step S130, based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lag, and electricity consumption is constructed by identifying equipment start-up and shutdown events. 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 electricity consumption patterns.

[0048] In step S140, random masking processing is performed on the temporal 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 anomalies are identified by joint contrastive learning, and feature correction data is output. In step S150, the feature correction data is input into the long short-term memory prediction network, and the confidence interval is dynamically adjusted using the quantile loss function. When the confidence interval is continuously exceeded, a feedback signal stream is generated.

[0049] In step S160, the feedback signal stream is parsed, and the weights of the long short-term memory prediction network are updated using a parameterized noise exploration strategy to achieve closed-loop optimization of the long short-term memory prediction network.

[0050] In step S170, the process ends.

[0051] The power load forecasting device proposed in Embodiment 2 of this invention achieves refined extraction of load characteristics 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. A quantile loss function is introduced to dynamically adjust the confidence level, and an online closed-loop optimization of the forecasting model parameters is achieved through a deep deterministic strategy gradient framework. This significantly improves the accuracy of power load forecasting, especially enhancing its dynamic adaptability to load fluctuations and effectively addressing the time-varying characteristics of electricity consumption patterns.

[0052] The power load forecasting device proposed in Embodiment 2 of this invention utilizes dynamic time warping matching and spectral clustering techniques to achieve equipment start-up and shutdown event identification and user homogeneous power 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 power consumption samples. This not only helps to uncover patterns in user power consumption behavior but also enables the timely detection of abnormal situations such as equipment failures, providing a reliable basis for the refined management of the power system.

[0053] The power load forecasting device proposed in Embodiment 2 of this invention improves the quality and stability of input features by using the NSGA-II multi-objective optimizer to perform evolutionary optimization of feature weights and combining data repair and anomaly detection with a dual-verification correction module. In the reinforcement learning framework, prediction error is used as the reward signal, and the collaborative work of the Actor-Critic network enables adaptive adjustment of model parameters, forming a closed-loop mechanism of "prediction-feedback-optimization," which significantly enhances the model's robustness and generalization ability in complex power consumption environments.

[0054] The power load forecasting device proposed in Embodiment 2 of this invention provides high-precision load forecasting results, which can offer scientific decision support to power dispatching departments. This helps to rationally plan power generation schedules, optimize grid operation, and reduce dispatching pressure caused by peak-valley differences. Simultaneously, in-depth analysis of user electricity consumption patterns can provide a basis for demand-side response strategy formulation, promote rational electricity use by users, improve energy efficiency, and drive the power system towards a more economical and environmentally friendly direction.

[0055] It should be noted that the present invention also provides an electronic device, including: a communication interface capable of interacting with other devices such as network devices; and a processor connected to the communication interface to enable information interaction with other devices, used to execute a power load forecasting method provided by one or more of the above technical solutions when running a computer program, wherein the computer program is stored in a memory. In practical applications, the various components of the electronic device are coupled together through a bus system. It is understood that the bus system is used to realize the connection and communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus. The memory in the embodiments of this application is used to store various types of data to support the operation of the electronic device. Examples of this data include any computer program used to operate on the electronic device. It is understood that the memory can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or magnetic tape storage. Volatile memory can be random access memory (RAM), which is used as an external cache.By way of example, but not limitation, many forms of RAM are available, 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 (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM). The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memory. The methods disclosed in the embodiments of this application can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by integrated logic circuits in the processor hardware or by instructions in software form. The processor 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 devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. A general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in a storage medium, which is located in memory. The processor reads the program from the memory and, in conjunction with its hardware, completes the steps of the aforementioned method. When the processor executes the program, it implements the corresponding processes in the various methods of the embodiments of this application; for simplicity, these will not be elaborated further here.

[0056] The description of the relevant parts of the power load forecasting device provided in Embodiment 2 of this application can be found in the detailed description of the corresponding parts of the power load forecasting method provided in Embodiment 1 of this application, and will not be repeated here.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0058] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting electricity load, characterized in that, Includes the following steps: Acquire current waveform data of the power system and extract the fundamental and harmonic components from the current waveform data; then perform geometric topological analysis of the spatial morphology of the current waveform to obtain the real-time load characteristic sequence. The real-time load characteristic sequence is segmented using a sliding time window; and the effective current value sequence within each time window is converted into a time-frequency domain energy distribution feature vector; based on the frequency domain energy distribution feature vector, a three-level feature library is constructed that integrates device fingerprint features, user behavior features, and environmental response features; Based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lag, and electricity consumption is constructed by identifying equipment start-up and shutdown events. 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 electricity consumption patterns. Random masking is performed on the temporal data of feature clusters to generate a mask sequence. The mask sequence is then input into the Transformer encoder to reconstruct the masked data. Anomalies are identified through joint contrastive learning, and feature correction data is output. The feature-corrected data is input into a long short-term memory 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 flow is analyzed, and a parameterized noise exploration strategy is used to update the weights of the long short-term memory prediction network, thereby achieving closed-loop optimization of the long short-term memory prediction network.

2. The power load forecasting method according to claim 1, characterized in that, The real-time load characteristic sequence is used to characterize the real-time data stream of instantaneous load status, including the effective value of voltage, the effective value of current, active power, reactive power, harmonic distortion rate, and crest factor.

3. The power load forecasting method according to claim 1, characterized in that, The real-time load characteristic sequence is segmented using a sliding time window; and the effective current value sequence within each time window is converted into a time-frequency domain energy distribution feature vector, specifically: The real-time load characteristic sequence is segmented using a fixed-duration sliding window, and the window length covers the complete start-stop cycle of the electrical equipment. The effective current value sequence within each time window is subjected to 6-level wavelet packet decomposition. The 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 domain energy distribution feature vector.

4. The power load forecasting method according to claim 1, characterized in that, A three-level feature library is constructed based on frequency domain energy distribution feature vectors, fusing device fingerprint features, user behavior features, and environmental response features. Specifically: Calculate the Pearson correlation coefficient matrix between the energy distribution of each frequency band and the equipment type, and select the associated frequency bands with a Pearson correlation coefficient greater than the threshold as the equipment fingerprint features; Identify high-frequency energy fluctuation patterns at a preset time scale in the frequency domain energy distribution feature vector, and construct user behavior features by combining them with the active power change rate; The environmental response characteristics are obtained by calculating the hysteresis cross-correlation function between the energy change of the corresponding characteristic frequency band of the temperature-sensitive device and the outdoor temperature.

5. The power load forecasting method according to claim 1, characterized in that, Based on the three-level feature library, a three-dimensional tensor model of meteorological factors, time lag, and electricity consumption is constructed by identifying equipment start-up and shutdown events. The three-dimensional tensor is then 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 electricity consumption patterns. Specifically: By calculating the similarity between the load curve and the equipment start-up and shutdown template, the equipment start-up and shutdown events and the lag response time of environmental factors are identified, and a three-dimensional tensor correlation model between meteorological factors and electricity consumption is generated. The three-dimensional tensor is input into the NSGA-II multi-objective optimizer. Using the silhouette coefficient and the Davidson-Bolding index as fitness functions, weighted evolution is performed on three types of features: temperature sensitivity, load fluctuation period, and power mutation amplitude. The Pareto front optimal feature weight combination is output. The weighted feature vector is input into the spectral clustering engine, and user feature clustering is performed based on the weighted Mahalanobis distance. After filtering out abnormal samples, feature clusters with homogeneous electricity consumption patterns are output.

6. The power load forecasting method according to claim 5, characterized in that, The equipment start-stop events include air conditioning compressor start-stop events; the environmental factor lag response time is the temperature lag response time constant; the spectral clustering engine determines the cluster center through Laplacian matrix eigenvalue decomposition, and outlier samples are those located more than 3 standard deviations from the cluster center.

7. The power load forecasting method according to claim 1, characterized in that, Random masking is performed on the temporal data of feature clusters to generate a mask sequence. The mask sequence is then input into the Transformer encoder to reconstruct the masked data. Combined with contrastive learning, anomalies are identified, and feature-corrected data is output. Specifically: Random masking processing is performed on the time-series data of the input feature clusters, and a mask sequence is generated by randomly selecting continuous time segments and setting them to zero according to a preset ratio. The mask sequence is input into the Transformer encoder, and the contextual dependencies are learned through a multi-head attention layer. Positional encoding is used to preserve temporal information and reconstruct the data of the masked segment. When the data of the reconstructed masked segment is valid, the feature space of the normal power consumption segment and the faulty power consumption segment is negatively sampled, the InfoNCE loss function is used to optimize the boundary of the abnormal mode discrimination, and the feature-corrected data is output after dual verification of sequence reconstruction and abnormal discrimination.

8. The power load forecasting method according to claim 1, characterized in that, The feature-corrected data is input into a long short-term memory 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. Specifically: A dual-channel long short-term memory prediction network is constructed, in which the main channel is used to input feature correction data and the auxiliary channel is used to input temperature sensitivity features. The feature-corrected data is input into the constructed dual-channel long short-term memory prediction network; the quantile loss function is used to dynamically calculate the prediction interval coverage probability, and the confidence level is adaptively adjusted according to the time-sharing load fluctuation characteristics; 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 direction and magnitude of the deviation.

9. The power load forecasting method according to claim 1, characterized in that, The feedback signal stream is analyzed, and a parameterized noise exploration strategy is used to update the weights of the Long Short-Term Memory (LSTM) prediction network, thereby achieving closed-loop optimization of the LTM prediction network. Specifically: By analyzing the direction and magnitude of the deviation in the feedback signal stream using an action network, a weight adjustment strategy for the hidden layer of a Long Short-Term Memory Prediction Network is generated. The strategy value function is evaluated through a network of critics, with the negative logarithm of the prediction error as the immediate reward; An experience replay mechanism is used to update the parameters of the long short-term memory prediction network, thereby achieving online closed-loop adaptive optimization of the network parameters.

10. A power load forecasting device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 9.

Images