Power time series data prediction method and device, electronic equipment and storage medium

Abstract

The application provides a power time series data prediction method and device, electronic equipment and storage medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: determining a residual signal of power time series data based on a physical baseline signal of the power time series data; inputting the residual signal into a residual prediction model to obtain a predicted residual signal; evaluating the predicted residual signal according to an effectiveness determination mechanism, and deciding to generate a prediction result of the power time series data based on the predicted residual signal and the physical baseline signal or only based on the physical baseline signal according to the evaluation result. The application decouples the power time series data into a physical baseline signal and a residual for separate processing, and uses an adaptive three-section gating mechanism to screen the effectiveness of the predicted residual signal, so that the key signal peak can be accurately captured on the basis of retaining the physical law constraint to ensure the safety of the power grid, the real load fluctuation and the sensor noise can be effectively distinguished, and the error scheduling instruction caused by misjudgment of the noise can be avoided.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for predicting power time-series data. Background Technology

[0002] Electricity time-series data forecasting refers to the technology of predicting future grid load data and electricity price fluctuations based on historical observation data. It is an important part of grid security dispatch and electricity market transactions.

[0003] Currently, multilayer perceptrons or recurrent neural networks are typically used to perform end-to-end modeling and training of power grid time series data.

[0004] However, this method lacks effective physical constraints, is prone to losing key signal features, and cannot capture millisecond-level electricity price spikes. Furthermore, when encountering outliers in the model output, it is difficult to distinguish whether the outlier is a real load fluctuation or sensor noise, which may lead to issuing incorrect power dispatch instructions and pose safety hazards. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for predicting power time-series data, which addresses the shortcomings of existing technologies such as lack of effective physical constraints, easy loss of key signal features, and difficulty in distinguishing between real load fluctuations and sensor noise, thus causing safety hazards. It achieves structural decoupling of physical laws and data fluctuations, and accurately captures key signal features while ensuring power grid safety.

[0006] This invention provides a method for predicting power time-series data, comprising the following steps: Based on the physical baseline signal of the power time series data, determine the residual signal of the power time series data; The residual signal is input into the residual prediction model to obtain the predicted residual signal output by the residual prediction model; Based on the predicted residual signal, the predicted residual signal is evaluated according to the validity determination mechanism to obtain the evaluation result. Based on the evaluation result, a decision is made to generate a prediction result for the power time series data based on the predicted residual signal and the physical baseline signal, or based solely on the physical baseline signal.

[0007] According to a power time-series data prediction method provided by the present invention, the physical baseline signal is determined based on the following steps: Construct the Hankel matrix of the power time-series data; Perform singular value decomposition on the Hankel matrix to obtain the left singular vector matrix, the central singular value matrix, and the right singular vector matrix; All singular values ​​of the central singular value matrix are sorted according to their numerical values ​​to obtain a singular value sequence, and a preset number of target singular values ​​are determined based on the singular value sequence. Based on the target singular value, the target left singular vector is determined from the left singular vector matrix, and the target right singular vector is determined from the right singular vector matrix; Based on the target singular values, the target left singular vector, and the target right singular vector, a low-rank matrix is ​​obtained; The low-rank matrix is ​​reconstructed to obtain the physical baseline signal.

[0008] According to a power time-series data prediction method provided by the present invention, constructing the Hankel matrix of the power time-series data includes: The power time series data is scrolled and truncated based on a sliding window of preset length to obtain multiple continuous local time series subsequences; The Hankel matrix is ​​constructed based on the multiple consecutive local temporal subsequences, wherein the elements in the Hankel matrix remain consistent along the anti-diagonal direction.

[0009] According to a power time-series data prediction method provided by the present invention, the physical baseline signal is determined based on the following steps: Determine the wavelet basis functions and the target decomposition level; Based on the wavelet basis function, multi-scale discrete wavelet decomposition is performed on the power time series data to obtain an approximate set of wavelet coefficients located at the target decomposition level; The physical feature components are determined based on the set of approximate wavelet coefficients for the target decomposition level; The physical feature components are reconstructed by inverse wavelet transform to obtain the physical baseline signal.

[0010] According to a power time-series data prediction method provided by the present invention, the physical baseline signal is determined based on the following steps: Empirical mode decomposition is performed on the power time series data to obtain residual components; The physical baseline signal is determined based on the residual components.

[0011] According to a power time-series data prediction method provided by the present invention, the step of inputting the residual signal into a residual prediction model and obtaining the prediction residual signal output by the residual prediction model includes: An initial weight matrix is ​​constructed based on the weights of all model parameters of the residual prediction model; Calculate the spectral norm of the initial weight matrix; The initial weight matrix is ​​normalized using the spectral norm to obtain a normalized weight matrix; The predicted residual signal is generated based on the normalized weight matrix and the residual signal.

[0012] According to a power time-series data prediction method provided by the present invention, the method includes: evaluating the prediction residual signal based on the prediction residual signal according to a validity determination mechanism to obtain an evaluation result; and, based on the evaluation result, deciding whether to generate a prediction result for the power time-series data based on the prediction residual signal and the physical baseline signal, or based solely on the physical baseline signal. The method further includes: Based on the historical statistical characteristics within the sliding window of the power time-series data, a first noise tolerance threshold and a second safety saturation threshold are calculated, and the second safety saturation threshold is greater than the first noise tolerance threshold. Obtain the amplitude of the predicted residual signal, and perform the following decision logic based on the amplitude: If the amplitude is less than or equal to the first noise tolerance threshold, the predicted residual signal is determined to be invalid background noise, and the prediction result for the power time series data is generated only based on the physical baseline signal. If the amplitude is greater than the first noise tolerance threshold and less than the second safety saturation threshold, the predicted residual signal is determined to be a valid fluctuation signal, and the physical baseline signal and the predicted residual signal are weighted and fused based on the dynamic weighting coefficient to generate a prediction result for the power time series data; wherein, the dynamic weighting coefficient is determined based on the mapping position of the amplitude within the numerical range formed by the first noise tolerance threshold and the second safety saturation threshold; If the amplitude is greater than or equal to the second safety saturation threshold, the amplitude of the predicted residual signal is limited to obtain a safety residual signal, and a prediction result for the power time series data is generated based on the physical baseline signal and the safety residual signal.

[0013] According to the present invention, a method for predicting power time-series data includes a step of weighting and fusing the physical baseline signal and the prediction residual signal based on dynamic weighting coefficients to generate a prediction result for the power time-series data, comprising: Obtain the time-series index of the predicted residual signal; Based on the time-series index, the target baseline value of the physical baseline signal is determined; The predicted residual signal and the target baseline value are weighted and fused based on the dynamic weighting coefficients to obtain the prediction result of the power time series data.

[0014] According to the power time series data prediction method provided by the present invention, the training process of the residual prediction model includes: Obtain historical power time-series data; The historical power time series data is decomposed to obtain the historical physical baseline signal, and the historical residual sample is determined based on the difference between the historical power time series data and the historical physical baseline signal. Determine the residual labels corresponding to the historical residual samples; Freeze the generation parameters of the historical physical baseline signal and iteratively execute the training process until the preset termination condition is met; The training process includes: Input the historical residual samples into the residual prediction model to obtain the predicted residuals output by the residual prediction model; Calculate the loss function value between the predicted residual and the residual label; Based on the loss function value, the model parameters of the residual prediction model are updated using the backpropagation algorithm.

[0015] The present invention also provides a power time series data prediction device, comprising the following modules: The first computing unit is used to acquire the physical baseline signal of power time-series data; A storage unit, connected to the first computing unit, is used to store the generation parameters of the physical baseline signal; The second calculation unit is used to determine the residual signal of the power time series data based on the physical baseline signal of the power time series data; input the residual signal into the residual prediction model, and obtain the prediction residual signal output by the residual prediction model; A physical isolation zone is used to block signal interaction between the first computing unit and the second computing unit; A gating unit, connected to the output of the first computing unit and the output of the second computing unit, is used to evaluate the prediction residual signal based on the prediction residual signal according to the validity determination mechanism, obtain the evaluation result, and, based on the evaluation result, decide whether to generate a prediction result for the power time series data based on the prediction residual signal and the physical baseline signal, or based solely on the physical baseline signal.

[0016] The present invention also provides a power time series data prediction device, which is implemented by a field-programmable gate array or an application-specific integrated circuit; The first computing unit and the second computing unit are respectively configured in different hardware logic partitions of the field-programmable gate array or the application-specific integrated circuit; The physical isolation strip is used to cut off the signal transmission path between the different hardware logical partitions.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the power time-series data prediction method as described above.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the power time-series data prediction method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the power time-series data prediction method as described above.

[0020] The present invention provides a power time-series data prediction method, device, electronic device, and storage medium that decouples power time-series data into physical baseline signals and residuals for separate processing. Based on the historical statistical characteristics of the data, a dynamic dual threshold is constructed, and a three-stage decision-making process of noise shielding, weighted fusion, or safety circuit breaker truncation is performed on the predicted residuals. Furthermore, an adaptive three-stage gating mechanism is used to effectively screen the predicted residual signals. This allows for the accurate capture of key signal spikes while preserving physical constraints to ensure power grid safety, effectively distinguishing between real load fluctuations and sensor noise, and avoiding the issuance of erroneous dispatch instructions due to misjudgment of noise. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts of the power time series data prediction method provided by the present invention.

[0023] Figure 2 This is one of the flowcharts for signal decomposition of power time-series data provided by the present invention.

[0024] Figure 3 This is the second flowchart illustrating the signal decomposition of power time-series data provided by the present invention.

[0025] Figure 4 This is a schematic diagram of the process for obtaining the prediction residual signal provided by the present invention.

[0026] Figure 5This is a schematic diagram of the process for generating power time series data prediction results based on predicted residual signals and physical baseline signals, provided by the present invention.

[0027] Figure 6 This is a schematic diagram of the training process of the residual prediction model provided by the present invention.

[0028] Figure 7 This is a hardware logic architecture diagram provided by the present invention.

[0029] Figure 8 This is the second flowchart of the power time-series data prediction method provided by the present invention.

[0030] Figure 9 This is a diagram showing the robust decomposition effect after introducing spectral norm constraints as provided by the present invention.

[0031] Figure 10 This is a diagram showing the effect of the benchmark control group provided by this invention without the introduction of spectral norm constraints.

[0032] Figure 11 This is a schematic diagram of the power time series data prediction device provided by the present invention.

[0033] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0035] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, 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.

[0036] The terms "upper," "lower," etc., indicating orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two elements. For those skilled in the art, the specific meaning of the above terms in the invention can be understood according to the specific circumstances.

[0037] The terms "first," "second," etc., used in this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more.

[0038] To facilitate a full understanding of the technical solution of this application, the following content is hereby introduced: With the development of smart grids, power load forecasting and system stability control have become core issues in energy management. Currently, existing solutions mainly fall into two categories: traditional statistical methods, such as Auto-Regressive Integrated Moving Average (ARIMA) and Kalman filtering; and data-driven methods based on deep learning, such as Back Propagation (BP) neural networks and Long Short-Term Memory (LSTM) networks.

[0039] Traditional deep learning solutions typically utilize multilayer perceptrons (MLPs) or recurrent neural networks (RNNs) to perform end-to-end modeling and training on power grid time-series data. By inputting historical data, they directly output predicted load values ​​or control commands. This approach often relies on massive datasets for training and utilizes graphics processing units (GPU) clusters for black-box prediction. Furthermore, traditional model preprocessing methods typically employ simple Gaussian filtering, wavelet transforms, or mean filtering to handle noise in power grid data, attempting to smooth data fluctuations. This approach is widely used in power grid dispatching.

[0040] However, existing technologies face a dilemma when it comes to balancing the need for both grid security and stability with the need for savvy market arbitrage opportunities. This dilemma manifests itself in the following ways: First, it suffers from poor noise resistance and impure feature extraction. Existing technologies typically assume that the noise is Gaussian white noise, but the noise in actual power grid operation is often nonlinear and complex. Simple filtering methods easily lose key signal features, and traditional physical models are too smooth to capture millisecond-level electricity price spikes, leading to missed high-frequency arbitrage opportunities in electricity trading.

[0041] Second, the model lacks stability and is prone to gradient explosion. When training with ordinary multilayer perceptrons or recurrent neural networks, existing neural networks lack effective constraints on the Lipschitz constant. When encountering adversarial samples or extreme data fluctuations, the network is prone to gradient explosion or gradient vanishing, causing the model training to diverge and failing to guarantee the robustness of the system under harsh conditions.

[0042] Third, there is a lack of physical safety fallback mechanisms. Existing pure algorithm models are usually black boxes. Once the algorithm outputs abnormal values, such as control commands exceeding the safe range, the system lacks an independent physical blocking mechanism. It is very easy to misinterpret random noise from sensors as sudden changes in grid load, leading to incorrect dispatch commands, causing grid frequency oscillations, and threatening energy security.

[0043] Fourth, it consumes a lot of computing power and is difficult to deploy. Existing improved models have high computational complexity and rely on expensive graphics processing units (GPUs), making them unsuitable for deployment on a large number of low-power smart meters or domestically produced field-programmable gate array (FPGA) edge terminals.

[0044] Therefore, the present invention provides a method, device, electronic device and storage medium for predicting power time series data, which realizes the structural decoupling of physical laws and data fluctuations, and accurately captures key signal features while ensuring power grid safety.

[0045] The following is combined with Figures 1-12 This invention describes the power time-series data prediction method, apparatus, electronic device, and storage medium provided by the present invention.

[0046] Figure 1 This is one of the flowcharts illustrating the power time-series data prediction method provided by the present invention, such as... Figure 1 As shown, the execution subject of the power time series data prediction method provided by the present invention can be an edge computing terminal, an embedded device, a smart meter, or a computer capable of executing the method of the present invention, etc. Unless otherwise specified, the following embodiments will be described using an edge computing terminal as an example.

[0047] As an optional embodiment, this power time-series data prediction method mainly includes, but is not limited to, the following steps: Step 110: Determine the residual signal of the power time series data based on the physical baseline signal of the power time series data.

[0048] Electricity time-series data refers to observational data collected in the power system in chronological order, reflecting the evolution of the power grid's operating status or market trading conditions over time. For example, electricity time-series data can be real-time data collected by smart meters, sensor terminals, or edge computing devices, including grid load data, node voltage data, and current data, or it can be real-time electricity price data from the electricity spot market.

[0049] The physical baseline signal can be obtained from power time series data through signal decomposition. Signal decomposition refers to the process of using mathematical transformations to break down the original observation data, which contains complex components, into different components to separate the main trend components and high-frequency fluctuation components. For example, signal processing algorithms such as singular value decomposition, wavelet transform, or empirical mode decomposition can be used to decompose the original power time series data into low-rank components representing steady-state trends and residual components representing transient fluctuations.

[0050] The physical baseline signal refers to the steady-state trend component extracted from power time-series data that conforms to the physical operating laws of the power grid. It represents the basic operating trajectory of the power grid under normal operating conditions and typically exhibits low-rank, periodic, and smooth characteristics. For example, the physical baseline signal can be a smooth curve reflecting the daily periodic or seasonal changes in power grid load, or a power grid voltage reference waveform after removing high-frequency noise and sudden disturbances. This physical baseline signal serves as the output of a rigid physical channel to ensure that the prediction results meet the basic requirements for power grid security.

[0051] The residual signal refers to the nonlinear fluctuation component remaining in power time-series data after removing the physical baseline signal. It contains information about the high-frequency variation characteristics of the power grid, sudden disturbances, and potential peak electricity market prices. For example, the residual can be calculated by subtracting the corresponding physical baseline signal from the original power time-series data. This process aims to separate the steady-state component, which contains physical laws, from the transient component, which contains randomness or abrupt changes.

[0052] Step 120: Input the residual signal into the residual prediction model and obtain the predicted residual signal output by the residual prediction model.

[0053] A residual prediction model refers to a neural network model used to fit and predict residuals. For example, a residual prediction model can be a lightweight multilayer perceptron, a one-dimensional convolutional neural network (1D-CNN), or a temporal convolutional network (TCN).

[0054] The predicted residual signal refers to the predicted residual value for future moments calculated by the residual prediction model based on the input historical residual data. For example, the predicted residual signal can represent the degree of deviation of the grid load or electricity price from the physical baseline at the next moment. This predicted residual signal reflects the residual prediction model's keen ability to capture potential arbitrage opportunities or sudden load changes.

[0055] Step 130: Based on the predicted residual signal, evaluate the predicted residual signal according to the validity judgment mechanism to obtain the evaluation result, and based on the evaluation result, decide whether to generate the prediction result of the power time series data based on the predicted residual signal and the physical baseline signal, or based only on the physical baseline signal.

[0056] The validity determination mechanism refers to a logical rule that adaptively adjusts the decision boundary based on the statistical characteristics of the signal. It aims to dynamically divide the invalid signal area, the valid signal area, and the dangerous signal area according to the current environmental noise level and safety tolerance of the power system, thereby realizing intelligent hierarchical processing of the prediction residual.

[0057] The prediction result for power time-series data refers to the final output predicted value used to guide grid dispatching or market transactions. For example, the amplitude of the predicted residual signal is compared with a preset threshold, and the result of the comparison determines whether to generate the prediction result for the power time-series data based on the predicted residual signal and the physical baseline signal or only based on the physical baseline signal.

[0058] Considering the presence of a large amount of small random noise in actual power grid operation, and the fact that traditional end-to-end models are prone to illusions due to overfitting noise, leading to erroneous scheduling instructions, this invention introduces a gating mechanism similar to a hardware dead-zone fuse. It only allows the correction of the physical baseline when the residual fluctuation is large enough and has practical physical significance or economic value. This effectively filters out false noise from sensors and achieves accurate capture of high-value abrupt change signals while ensuring the safety of power grid frequency and the preservation of physical laws.

[0059] The power time series data prediction method provided by this invention decouples power time series data into physical baseline signals and residuals for separate processing, and uses an adaptive three-stage gating mechanism to effectively screen the prediction residual signals. This allows for the accurate capture of key signal spikes while preserving physical constraints to ensure power grid safety, effectively distinguishing between real load fluctuations and sensor noise, and avoiding the issuance of erroneous dispatch instructions due to misjudgment of noise.

[0060] Figure 2 This is one of the flowcharts for signal decomposition of power time-series data provided by the present invention, such as... Figure 2 As shown, as another optional embodiment provided by the present invention, the physical baseline signal is determined based on the following steps: Step 210: Construct the Hankel matrix of the power time series data.

[0061] A Hankel matrix is ​​a matrix with constant antidiagonal elements that can embed one-dimensional time series data into a high-dimensional space. For example, by setting a sliding window length, a Hankel matrix can arrange power time series data in chronological order into a multi-row, multi-column format, such that the element in the i-th row and j-th column is the same as the element in the (i+1)-th row and (j-1)-th column, thus completing the mapping from one-dimensional time series data to a two-dimensional matrix.

[0062] Step 220: Perform singular value decomposition on the Hankel matrix to obtain the left singular vector matrix, the central singular value matrix, and the right singular vector matrix.

[0063] Singular value decomposition (SVD) is a matrix factorization method used to decompose a matrix into the product of three matrices, thereby extracting the matrix's characteristic patterns. For example, the Hankel matrix can be decomposed into the product of an orthogonal matrix (left singular vector matrix), a diagonal matrix (central singular value matrix), and a transposed orthogonal matrix (right singular vector matrix).

[0064] Step 230: Sort all the singular values ​​of the central singular value matrix according to their numerical values ​​to obtain a singular value sequence, and determine a preset number of target singular values ​​based on the singular value sequence.

[0065] Specifically, the elements on the diagonal of the central singular value matrix are arranged in descending order to form a singular value sequence that reflects the energy strength of each mode. For example, based on the energy decay characteristics of the singular value sequence, the first k largest singular values ​​can be selected as a preset number of target singular values. These larger singular values ​​usually correspond to the main trend components with a large energy proportion in the power grid data, such as daily cycles or seasonal trends, while smaller singular values ​​usually correspond to noise or high-frequency disturbances.

[0066] Step 240: Determine the target left singular vector from the left singular vector matrix and the target right singular vector from the right singular vector matrix based on the target singular values.

[0067] The target left singular vector refers to the column vector in the left singular vector matrix that corresponds one-to-one with a predetermined number of singular values ​​at their index positions. For example, if the first k singular values ​​are selected, then the first k columns of the left singular vector matrix are extracted as the target left singular vector.

[0068] The target right singular vector refers to the row or column vector in the right singular vector matrix that corresponds one-to-one with a predetermined number of singular values ​​at their index positions (depending on the transpose of the matrix definition). For example, the first k rows or the first k columns of the right singular vector matrix can be extracted as the target right singular vector.

[0069] Step 250: Based on the target singular values, the target left singular vector, and the target right singular vector, obtain the low-rank matrix.

[0070] Specifically, a new diagonal matrix is ​​constructed using a selected preset number of singular values, and then matrix multiplication is performed between this matrix and the matrix formed by the left singular vector of the target and the matrix formed by the right singular vector of the target.

[0071] For example, a new matrix that approximates the original Hankel matrix can be reconstructed by truncating singular value decomposition. A low-rank matrix refers to a matrix obtained through the above reconstruction with a rank less than that of the original Hankel matrix. It removes the small singular value components representing noise and high-frequency fluctuations, retaining only the low-rank structural features representing the main physical laws.

[0072] Step 260: Reconstruct the low-rank matrix to obtain the physical baseline signal.

[0073] Specifically, an inverse Hankel transformation is performed on the low-rank matrix, which involves averaging the elements on the same anti-diagonal line of the low-rank matrix to restore it to a one-dimensional time series. For example, the arithmetic mean of the anti-diagonal elements of the low-rank matrix is ​​used as the signal value at the corresponding time, thereby generating a smooth physical baseline signal that has been freed from high-frequency noise interference. This signal reflects the steady-state operating trajectory of the power system.

[0074] It should be noted that, considering that power time series data usually implicitly contain low-rank structural characteristics determined by the power consumption patterns and physical inertia of the power grid, that is, the main energy is concentrated in a few modes, while noise is often distributed in all modes and has relatively small energy, this invention adopts a singular value decomposition method based on the Hankel matrix. By retaining the main singular values ​​and reconstructing the low-rank matrix, the projection separation of the signal manifold is achieved. This allows the physical baseline signal representing the rigid operation law of the power grid to be extracted from complex observation data. While effectively filtering out random noise interference, the periodic trend and phase characteristics of the power grid load are preserved to the maximum extent, providing an absolutely safe physical reference benchmark for subsequent residual prediction.

[0075] The power time series data prediction method provided by this invention constructs the Hankel matrix of power time series data and performs singular value decomposition to reconstruct a low-rank matrix. It can utilize the low-rank manifold characteristics of power grid data to orthogonally separate the steady-state trend containing the main physical laws from random noise at the feature subspace level, thereby generating a high-fidelity, phase-lag-free physical baseline signal, ensuring the accuracy of subsequent residual calculation and the effective constraint of the system on physical laws.

[0076] In another embodiment of the present invention, constructing a Hankel matrix of power time series data includes: rolling and truncating power time series data based on a sliding window of a preset length to obtain multiple consecutive local time series subsequences; constructing a Hankel matrix based on the multiple consecutive local time series subsequences, wherein the elements in the Hankel matrix along the anti-diagonal direction remain consistent.

[0077] Specifically, a fixed-length sliding window is set. Starting from the beginning of the power time series data, the first local time series subsequence within the coverage area of ​​the window is extracted and used as the first column vector of the Hankel matrix. Then, the sliding window is moved backward along the time axis by one sampling step, and the second local time series subsequence is extracted and used as the second column vector of the Hankel matrix. This process is repeated until the sliding window moves to the end of the preset observation window, thereby arranging the one-dimensional power time series data into a two-dimensional matrix structure with multiple rows and columns.

[0078] For example, assuming the power time series data is a voltage sequence collected every second, and the sliding window length is set to 3, the first column of the constructed matrix consists of the data from the 1st to the 3rd second, and the second column consists of the data from the 2nd to the 4th second. At this time, the element in the 1st row and 2nd column of the matrix and the element in the 2nd row and 1st column are both the voltage values ​​of the 2nd second. This special arrangement ensures that all elements in the matrix that are parallel to the anti-diagonal are numerically identical.

[0079] It should be noted that the determination of the physical baseline signal and the generation of the prediction residual signal are strictly limited to the historical time scale covered by the sliding window, so as to ensure that the prediction process satisfies temporal causality and eliminates phase lag.

[0080] The power time series data prediction method provided by this invention constructs a Hankel matrix by using a sliding window scrolling method, which can embed one-dimensional time series data into a high-dimensional phase space. This not only preserves the time dependency of the original data, but also utilizes the consistent anti-diagonal elements of the Hankel matrix to provide a foundation for the accurate extraction of low-rank physical baselines through singular value decomposition.

[0081] Figure 3 This is the second flowchart illustrating the signal decomposition of power time-series data provided by this invention, as shown below. Figure 3 As shown, as another optional embodiment provided by the present invention, the physical baseline signal is determined based on the following steps: Step 310: Determine the wavelet basis function and the target decomposition level.

[0082] Wavelet basis functions refer to mathematical oscillatory waveforms with finite energy that decay rapidly in the time domain. As the basic unit of signal transformation, they determine the ability to extract local features of power data. For example, wavelet basis functions can be Dobesi wavelets, Haar wavelets, etc., and different basis functions are suitable for capturing different types of power grid waveform features.

[0083] The number of target decomposition layers refers to the depth of multi-scale decomposition of the signal, which determines the fineness of frequency separation. For example, the number of target decomposition layers can be set to 3, 5 or more layers depending on the sampling frequency and the required baseline smoothness. The deeper the layers, the lower the frequency of the extracted approximate components, and the closer they are to the DC or fundamental frequency trend.

[0084] Step 320: Based on wavelet basis functions, perform multi-scale discrete wavelet decomposition on the power time series data to obtain an approximate set of wavelet coefficients located at the target decomposition level.

[0085] Specifically, using the low-pass and high-pass filters corresponding to the selected wavelet basis functions, layer-by-layer convolution and downsampling operations are performed on the power time series data. For example, the original signal is decomposed into high-frequency detail coefficients and low-frequency approximation coefficients, and the approximation coefficients are further decomposed into the next layer until the preset target decomposition layer is reached, thereby obtaining the set of approximate wavelet coefficients at that level.

[0086] Step 330: Determine the physical feature components based on the set of approximate wavelet coefficients of the target decomposition level.

[0087] Specifically, approximate wavelet coefficients located at the target decomposition level are selected as physical feature components, or thresholding is applied to these approximate coefficients to further filter out residual noise. For example, low-frequency approximate coefficients containing the main energy and long-term trends of the power grid are retained, while high-frequency detail coefficients at each level are discarded, because high-frequency detail coefficients usually correspond to random noise or transient disturbances and do not fall within the scope of the physical baseline.

[0088] Step 340: Perform inverse wavelet transform on the physical feature components to reconstruct the physical baseline signal.

[0089] Specifically, wavelet reconstruction algorithms are used to upsample and deconvolve the selected physical feature components, restoring them from the transform domain back to the time domain. For example, by using single-branch reconstruction, a smooth time series is generated using only the lowest-level approximate wavelet coefficients; this series is the physical baseline signal with high-frequency noise removed.

[0090] It should be noted that wavelet transform has good time-frequency localization characteristics and can adaptively process non-stationary power time series data. By selecting deep approximation coefficients for reconstruction, it is equivalent to applying a physically meaningful low-pass filter to the signal, thereby effectively extracting the steady-state trend of power grid operation as a safety baseline, while leaving high-frequency fluctuations to be processed by the subsequent residual network, which is also in line with the design concept of soft and hard decoupling of this invention.

[0091] The power time series data prediction method provided by this invention utilizes multi-scale discrete wavelet decomposition technology to extract approximate wavelet coefficients located at the target decomposition level and reconstruct them through inverse transform. It can leverage the excellent time-frequency localization characteristics of wavelet transform, which is equivalent to implementing a low-pass filter with a clear physical cutoff frequency. This effectively separates the low-frequency steady-state trend in the power grid data as a physical baseline. While filtering out high-frequency random noise, it retains the physical characteristics of the signal at different scales, providing a smooth and stable reference benchmark for subsequent residual analysis.

[0092] In another embodiment of the present invention, the power time series data is decomposed to obtain a physical baseline signal, including: performing empirical mode decomposition on the power time series data to obtain residual components; and determining the physical baseline signal based on the residual components.

[0093] Specifically, Empirical Mode Decomposition (EMD) is an adaptive time-frequency data processing method that does not require pre-defined basis functions. Instead, based on the data's inherent time-scale characteristics, it iteratively decomposes complex power time-series data into several intrinsic mode function components with frequencies ranging from high to low, and a final residual component. For example, by continuously subtracting the local mean envelope from the original data until the remaining portion no longer satisfies the definition conditions of the intrinsic mode functions, the residual component is obtained. The residual component refers to the monotonic trend term or long-period trend term remaining in the signal after removing all oscillation modes, representing the inherent overall trend of the data.

[0094] For example, the residual component reflects the macroscopic trend of grid load changes over a day or a quarter. Determining the physical baseline signal based on the residual component means either directly using the residual component as the physical baseline signal, or superimposing the residual component with several low-frequency intrinsic mode function components to obtain the physical baseline signal.

[0095] It should be noted that, considering that the residual component is naturally smooth and monotonic and does not contain any high-frequency noise or instantaneous fluctuations, defining it directly as the physical baseline signal can accurately lock the steady-state equilibrium point of the power grid operation, which meets the physical channel's definition requirements for the safety baseline.

[0096] The power time series data prediction method provided by this invention extracts residual components by performing empirical mode decomposition on power time series data. It can utilize the data adaptive characteristics of the algorithm to automatically separate the intrinsic mode functions of each order in the signal without the need for preset basis functions. The monotonic trend term after stripping all oscillation modes is used as the physical baseline signal, thereby ensuring that the baseline signal can truly reflect the inherent overall trend of power grid operation and providing a pure physical reference that has removed all high-frequency and mid-frequency fluctuation interference for subsequent residual prediction.

[0097] Figure 4 This is a schematic diagram of the process for obtaining the predicted residual signal provided by the present invention, as shown below. Figure 4 As shown, as another optional embodiment provided by the present invention, the residual signal is input into the residual prediction model to obtain the predicted residual signal output by the residual prediction model, including but not limited to the following steps: Step 410: Construct an initial weight matrix based on the weights of all model parameters of the residual prediction model.

[0098] The initial weight matrix refers to the set of parameters connecting neurons in adjacent layers of a neural network. It determines the transformation intensity and mapping relationship of the input signal as it is transmitted between network layers. For example, in a multilayer perceptron or convolutional neural network used for residual prediction, the initial weight matrix can be the convolutional kernel parameter matrix or the weight matrix of the fully connected layer that has not yet been normalized in the current training iteration.

[0099] Step 420: Calculate the spectral norm of the initial weight matrix.

[0100] Specifically, singular value decomposition or power iteration algorithms are used to approximate the initial weight matrix to obtain the largest singular value. For example, the spectral norm is mathematically defined as the maximum stretching factor of a matrix over a vector. Its value directly corresponds to the Lipschitz constant of the network layer, reflecting the sensitivity of the model output to small perturbations in the input or the maximum amplification factor of the gradient.

[0101] Step 430: Normalize the initial weight matrix using the spectral norm to obtain the normalized weight matrix.

[0102] Specifically, each element in the initial weight matrix is ​​divided by the calculated spectral norm, so that the spectral norm of the processed matrix equals 1. For example, this global scaling operation forces the weight matrix of each layer of the neural network to satisfy the spectral normalization condition, thereby restricting the mapping function of the network layer to the constraint of Lipshitz continuity, that is, limiting the maximum amplification factor of the network to input perturbations to 1, ensuring that the model does not amplify small noise in the input data.

[0103] Step 440: Generate the prediction residual signal based on the normalized weight matrix and the residual signal.

[0104] Specifically, a normalized weight matrix is ​​used to replace the original weight matrix to perform a linear transformation and subsequent nonlinear activation operation on the input residual data. For example, during the inference or forward propagation process of the residual prediction model, a constrained normalized weight matrix is ​​used to calculate the predicted residual value at the next time step, thereby outputting the predicted residual signal.

[0105] It should be noted that, considering that ordinary neural network models lack effective constraints on the Lipshitz constant, they are prone to gradient explosion or gradient vanishing when faced with extreme fluctuations or adversarial noise in power grid data, leading to model training divergence or unstable output predictions. Therefore, this invention introduces a spectral normalization mechanism to constrain the global gain of the model by limiting the spectral norm of the weight matrix, thereby preventing the model from overfitting to noise and ensuring the numerical stability and robustness of the residual prediction model under harsh operating conditions.

[0106] The power time series data prediction method provided by this invention can effectively constrain the Lipshitz constant of the residual prediction model by calculating the spectral norm of the initial weight matrix and normalizing it, thus preventing training divergence caused by gradient explosion. This significantly improves the robustness and stability of the model in the face of nonlinear noise and extreme data fluctuations while ensuring that the model can keenly capture high-frequency residuals.

[0107] In another embodiment of the present invention, based on the predicted residual signal, the predicted residual signal is evaluated according to a validity determination mechanism to obtain an evaluation result. Based on the evaluation result, a decision is made to generate a prediction result for the power time series data based on the predicted residual signal and the physical baseline signal, or based solely on the physical baseline signal, including: Based on the historical statistical characteristics within a sliding window of power time-series data, a first noise tolerance threshold and a second safety saturation threshold are calculated, with the second safety saturation threshold being greater than the first noise tolerance threshold.

[0108] Obtain the amplitude of the predicted residual signal, and perform the following decision logic based on the amplitude: If the amplitude is less than or equal to the first noise tolerance threshold, the predicted residual signal is determined to be invalid background noise, and a prediction result for the power time series data is generated only based on the physical baseline signal. If the amplitude is greater than the first noise tolerance threshold and less than the second safety saturation threshold, the predicted residual signal is determined to be a valid fluctuation signal, and the physical baseline signal and the predicted residual signal are weighted and fused based on dynamic weighting coefficients to generate a prediction result for the power time series data. The dynamic weighting coefficients are determined based on the mapping position of the amplitude within the numerical range formed by the first noise tolerance threshold and the second safety saturation threshold. If the amplitude is greater than or equal to the second safety saturation threshold, the amplitude of the predicted residual signal is limited to obtain a safety residual signal, and a prediction result for the power time series data is generated based on the physical baseline signal and the safety residual signal.

[0109] Specifically, a fixed-length historical observation window is first obtained, and statistical indicators of the data within the window are calculated in real time, such as the mean and standard deviation. Typically, the first noise tolerance threshold is set as the mean plus three times the standard deviation to cover 99.7% of the normal Gaussian white noise range; while the second safety saturation threshold is set as several times the first noise tolerance threshold, or the maximum instantaneous fluctuation limit allowed by the power grid physical equipment.

[0110] Once the predicted residual signal is generated at the next time step, its amplitude is compared with two dynamically changing thresholds. For valid fluctuations between the two thresholds, linear interpolation or a nonlinear activation function (such as the sigmoid function) is used to map the amplitude to a weighting coefficient between 0 and 1. This means that the closer the residual is to the second safe saturation threshold, the larger its weighting coefficient, and the greater the proportion superimposed on the physical baseline signal. For abnormally large signals exceeding the second safe saturation threshold, a hard clipping operation is performed, that is, the residual value is directly truncated to the size of the second safe saturation threshold, or soft saturation compression is performed, such as using a hyperbolic tangent function to compress it within the safe boundary, thus obtaining a safe residual signal.

[0111] For example, assuming that based on the electricity price fluctuation data of the past hour, the noise tolerance threshold is calculated to be 5 yuan / MWh, and the maximum allowable fluctuation threshold for equipment safety is 50 yuan / MWh. If the model predicts that the electricity price fluctuation in the next moment will be 3 yuan / MWh, which is less than the noise tolerance, the system considers this a meaningless random disturbance and directly outputs a smooth physical baseline price. If the predicted fluctuation is 25 yuan / MWh, which is between the two thresholds, the system calculates a weighting coefficient of approximately 0.44, and then adds the 11 yuan / MWh fluctuation to the baseline. If the predicted fluctuation suddenly reaches 100 yuan / MWh, far exceeding the safety threshold, it is determined that a data attack or sensor failure may have occurred. At this time, the residual is immediately limited to 50 yuan / MWh and added to the baseline for output. This not only prevents grid oscillations caused by erroneous instructions but also retains warning information about extreme market conditions to a certain extent. This dynamic three-stage processing mechanism achieves a smooth switch from conservative physical protection to aggressive market arbitrage and then to a safety circuit breaker.

[0112] Considering the highly non-stationary and uncertain operating environment of power systems, background noise levels dynamically change over time, and network attacks or sensor failures can trigger extreme abnormal data fluctuations at any time. Using fixed thresholds or a single superposition strategy makes it difficult to simultaneously ensure the security of power grid regulation and capture high-frequency trading opportunities in the market. Therefore, this invention proposes a dynamic dual-threshold, three-stage gating mechanism based on sliding window statistical characteristics. An adaptive silent dead zone is constructed using the first threshold to shield random noise, and a safety saturation boundary is constructed using the second threshold to circuit breaker against abnormal impacts. A dynamic weighted channel based on amplitude mapping is established between the two thresholds. This enables automatic adjustment of signal acquisition strategies under complex and variable operating conditions. It effectively filters out meaningless background clutter to maintain baseline stability, sensitively captures real load fluctuations to respond to market demands, and provides a mandatory physical safety boundary when encountering extreme interference. This achieves a deep integration and dynamic balance between data-driven prediction and power grid physical safety constraints.

[0113] The power time-series data prediction method provided by this invention constructs a dynamic dual-threshold system based on historical statistical features of a sliding window, and performs a three-stage adaptive decision-making logic on the prediction residual signal accordingly. It can automatically adjust the degree of acceptance and processing strategy of the residual signal according to changes in the real-time operating environment of the power grid. While effectively filtering out random interference and sensitively capturing real market fluctuations, it provides a mandatory physical security boundary for the system to prevent abnormal data caused by sensor failure or malicious attacks from impacting the power grid dispatching system. Thus, it achieves a dynamic balance and optimal coordination between prediction sensitivity, noise resistance robustness and physical security.

[0114] Figure 5This is a flowchart illustrating the process of generating power time-series data prediction results based on predicted residual signals and physical baseline signals, as provided by the present invention. Figure 5 As shown, as another optional embodiment provided by the present invention, the physical baseline signal and the prediction residual signal are weighted and fused based on dynamic weighting coefficients to generate a prediction result for power time series data, including but not limited to the following steps: Step 510: Obtain the time-series index of the predicted residual signal.

[0115] The time-series index of the predicted residual signal refers to the label or number that identifies the position of the predicted residual signal on the time axis. It is used to ensure strict alignment of the residual data with the physical baseline data in the time dimension. For example, the time-series index can be the timestamp of the prediction time or the relative time step number relative to the start point of the observation window.

[0116] Step 520: Determine the target baseline value of the physical baseline signal based on the time-series index.

[0117] The target baseline value refers to the specific amplitude of the physical baseline signal at the time corresponding to the time series index. It represents the steady-state physical laws or base load level that the power grid should follow at that prediction time. For example, the target baseline value can be obtained by searching or interpolating in the physical baseline sequence reconstructed after singular value decomposition, based on the time series index.

[0118] Step 530: The predicted residual signal and the target baseline value are weighted and fused based on the dynamic weighting coefficient to obtain the prediction result of the power time series data.

[0119] Specifically, the predicted residual signal values ​​are weighted and fused with the target baseline values ​​to synthesize the final predicted value. For example, by weighted fusion, the predicted residual signal representing high-frequency market fluctuations is superimposed on the target baseline value representing the grid security boundary, thereby generating a power time series data prediction result containing rich detailed features.

[0120] It should be noted that, considering that while physical baseline signals are safe and stable, they are prone to missing high-frequency arbitrage opportunities, and while predictive residual signals are sensitive, they lack physical constraints and direct use may bring security risks, this invention, after confirming the validity of the residuals through threshold judgment, adopts a weighted fusion method to generate the final result. This enables the organic integration of rigid physical laws and flexible market fluctuations, ensuring the basic safety of power grid dispatch and accurately capturing instantaneous price spikes in the electricity spot market.

[0121] The power time series data prediction method provided by this invention uses time series indexing to strictly align and superimpose the predicted residual signal with the physical baseline signal in time, which can ensure that the high-frequency fluctuation component accurately acts on the corresponding steady-state trend. This achieves accurate reconstruction of the rigid physical channel and the flexible data channel in the time domain, thus taking into account both the physical inertia of power grid operation and the instantaneous change characteristics of market transactions in the output results.

[0122] Figure 6 This is a schematic diagram of the training process of the residual prediction model provided by the present invention, as shown below. Figure 6 As shown, as another optional embodiment provided by the present invention, the training process of the residual prediction model includes, but is not limited to, the following steps: Step 610: Obtain historical power time series data.

[0123] Historical power time-series data refers to power grid observation records stored in a database over a past period. It includes known timestamps, load values, or electricity price values, and is used to construct training sets for supervised learning. For example, it can be obtained by retrieving historical log data of voltage, current, active power, and nodal marginal prices collected over the past year from power dispatch centers or distributed smart meters.

[0124] Step 620: Perform signal decomposition on the historical power time series data to obtain the historical physical baseline signal, and determine the historical residual sample based on the difference between the historical power time series data and the historical physical baseline signal.

[0125] Historical physical baseline signals refer to the steady-state trend components extracted from historical data using the same signal decomposition algorithms as those used in the inference phase (such as singular value decomposition). For example, a smooth curve obtained by constructing a Hankel matrix from historical data and performing low-rank reconstruction.

[0126] Historical residual samples refer to the fluctuation portion of historical data after removing the physical baseline, and they serve as input features for neural networks. For example, subtracting the historical physical baseline signal from historical power time-series data yields a series of numerical sequences reflecting high-frequency fluctuations.

[0127] Step 630: Determine the residual labels corresponding to the historical residual samples.

[0128] The residual label refers to the actual residual value used as the true value during training, at the next time step or in the future. For example, it can be calculated by subtracting the corresponding physical baseline value at the next time step from the actual observation value, and is used to guide the residual prediction model in calculating the prediction error.

[0129] Step 640: Freeze the generation parameters of the historical physical baseline signal and iteratively execute the training process until the preset termination condition is met.

[0130] Specifically, in a deep learning framework, the parameters used to generate the physical baseline signal (such as the singular vectors obtained from singular value decomposition) are set to be non-trainable, preventing them from participating in gradient descent updates. A preset termination condition refers to the criterion for stopping model training. For example, it could be reaching a preset upper limit for the number of training epochs, or the loss function value on the validation set no longer decreasing significantly.

[0131] The training process includes: inputting historical residual samples into the residual prediction model to obtain the predicted residuals output by the residual prediction model; calculating the loss function value between the predicted residuals and the residual labels; and updating the model parameters of the residual prediction model based on the loss function value using the backpropagation algorithm.

[0132] It should be noted that when updating the model parameters of the residual prediction model using the backpropagation algorithm, this invention introduces a gradient blocking mechanism, which forcibly cuts off the backpropagation path of the neural network to the physical baseline generation process. This means that the error of the residual prediction model is only used to update the model's own weights and will never modify the generation parameters of the physical baseline signal. This design ensures the independence of physical laws and prevents the neural network from interfering with the physical baseline signal by adjusting the baseline parameters in pursuit of fitting accuracy, thereby guaranteeing the purity and security of the physical channel.

[0133] The power time series data prediction method provided by this invention freezes the generation parameters of the physical baseline signal during the training process of the residual prediction model, thereby implementing gradient blocking at the algorithm level and cutting off the modification path of the physical baseline generation logic by the backpropagation of neural network error. This prevents the residual prediction model from destroying the steady-state characteristics of the physical baseline in order to overfit noise, ensuring that the independence of physical laws is not polluted by data-driven algorithms, and achieving complete decoupling of the physical channel and the algorithm channel during the training phase.

[0134] Figure 7 This is a hardware logic architecture diagram provided by the present invention, such as... Figure 7 As shown, this hardware architecture employs a dual-channel parallel pipeline design with hardware and software collaboration, typically embedded in domestically developed and controllable field-programmable gate arrays (FPGAs) or application-specific integrated circuit (ASIC) chip logic. The architecture mainly includes an input terminal, a safety baseline channel in the upper path, an arbitrage capture channel in the lower path, a physical isolation band in the middle, and an adder at the output terminal. Specifically, the input terminal receives real-time grid timing data from smart meters or sensors and distributes the data to the upper and lower channels.

[0135] The on-line safety baseline channel (first calculation channel) includes a read-only memory (ROM) and a multiplier array. The ROM serves as a storage unit to permanently store the physical characteristic basis parameters (SVD characteristic parameter library) after singular value decomposition. These parameters are considered as immutable physical constraints of the power grid. The multiplier array is connected to the ROM and configured to perform only linear matrix multiplication operations, thereby generating an absolutely safe physical baseline signal.

[0136] The arbitrage capture channel in the lower path includes a logic operation unit and a dead-time comparator. The logic operation unit acts as the inference engine of the quantized multilayer perceptron, used to perform fixed-point neural network inference to generate the prediction residual signal. The dead-time comparator is connected in series at the output of the logic operation unit, and it acts as a hardware-level dead-time fuse, thereby deciding whether to generate the prediction result for the power time series data based on the prediction residual signal and the physical baseline signal or only based on the physical baseline signal.

[0137] It is worth noting that a physical isolation strip is installed between the upper and lower channels. The two channels are physically isolated on the chip layout, with no signal interaction interface, forming a hardware-level gradient and signal physical isolation, fundamentally eliminating the possibility of artificial intelligence models interfering with the physical baseline. Finally, the adder connects the outputs of the upper and lower channels, superimposing the physical baseline signal with the filtered prediction residual signal to generate the final power time series data prediction result.

[0138] Figure 8 This is the second flowchart illustrating the power time-series data prediction method provided by this invention, as shown below. Figure 8 As shown, this process specifically demonstrates the algorithm execution logic based on a hardware-software co-engineered dual-channel architecture. First, step S1 acquires time-series data, i.e., obtains real-time observation sequences from the power grid terminal. Then, step S2 performs physical manifold projection to extract the main trend, using techniques such as singular value decomposition to extract the physical baseline signal representing the power grid's steady state. Here, a gradient blocking mechanism is introduced to freeze the generated parameters to prevent subsequent error backpropagation from interfering with the physical laws. Next, step S3 calculates the residuals to obtain the difference between the observed data and the main trend. Then, step S4 performs constrained nonlinear mapping, i.e., fitting the residuals using a residual prediction model (such as a multilayer perceptron) with spectral norm constraints, and preventing gradient explosion by limiting the Lipshitz constant of the model. Then, step S5 performs adaptive gating, i.e., by comparing the amplitude of the predicted residuals with a preset threshold, only retaining fluctuations with economic or physical significance. Finally, step S6 performs superposition output, linearly superimposing the effective predicted residuals with the physical baseline signal based on the gating result to generate the final prediction result.

[0139] Figure 9This is a robust decomposition effect diagram provided by the present invention after introducing spectral norm constraints, such as... Figure 9 As shown, four vertically arranged subplots illustrate the superior performance of the method of this invention in processing noisy power data. The vertical axis of the first subplot represents the electricity price ($ / MWh), showing the robust decomposition results with spectral norm normalization. The thin gray line represents the raw observation data containing a large amount of random noise, the red dashed line represents the ground truth, and the solid green line represents the reconstructed signal generated by this invention. The box in the upper left corner shows that the mean squared error (MSE) is only 1.41, indicating that the reconstructed signal effectively filters out background noise while closely following the trajectory of the true value.

[0140] The second subplot, with its vertical axis representing electricity price ($ / MWh), shows the low-frequency component obtained from the decomposition, i.e., the physical baseline signal extracted in this invention. The Beta MSE, marked in the upper left corner, is 0.38. The blue solid line represents the trend estimate, exhibiting smooth and stable periodic characteristics, unaffected by high-frequency noise. The third subplot, with its vertical axis representing the residual ($ / MWh), shows the sparse component obtained from the decomposition, i.e., the predicted residual signal. The upper left corner indicates a maximum amplitude (Max |a|) of 7.08 and a sparsity of 96.0%. The red curve represents sparse anomalies, and the orange dots mark outliers accurately identified by the model. These correspond to price spikes or load surges in the power grid that have arbitrage value. The fourth subplot shows the training loss curve, with the horizontal axis representing the training epoch and the vertical axis representing the loss value. It can be seen that the total loss, fidelity loss, and regularization term all decrease rapidly in the early stages of training and tend to converge and stabilize after about 50 epochs. This strongly proves that after introducing the spectral norm constraint, the model effectively avoids the gradient explosion problem and has extremely high numerical stability and convergence speed.

[0141] Figure 10 This is a diagram showing the effect of the benchmark control group provided by this invention without the introduction of spectral norm constraints, as shown in the figure. Figure 10 As shown, the decomposition results of the traditional end-to-end model are presented in the absence of spectral norm constraints (i.e., without restrictions on the Lipschitz constant), which are used to compare with... Figure 9This presents a stark contrast. The first subplot, with its vertical axis representing electricity price ($ / MWh), shows the standard decomposition results. The solid orange line represents the reconstructed signal, which exhibits violent oscillations and almost perfectly fits the random noise in the original observation data (the thin gray line). The mean square error (MSE) marked in the upper left corner is as high as 59.31, significantly deviating from the ground truth (the red dashed line).

[0142] The second subplot, with the ordinate representing electricity price ($ / MWh), shows the low-frequency components obtained from the decomposition. The BetaMSE, marked in the upper left corner, is also as high as 59.31. The trend estimate, represented by the blue solid line, exhibits a highly jagged and uneven pattern, failing to extract a stable physical baseline, indicating that the model cannot distinguish between physical laws and high-frequency noise. The third subplot, with the ordinate representing residuals ($ / MWh), shows the sparse components obtained from the decomposition. The upper left corner indicates that the maximum amplitude (Max |a|) is 0.00, and the sparsity is 100.0%. The red curve consistently hovers around 0, indicating that the model incorrectly classifies all anomalous fluctuations as low-frequency components, failing to identify any valid sparse outliers.

[0143] The training loss curve in the fourth subplot shows that the loss value on the ordinate drops rapidly to 10 within a very short number of training epochs. -9 This unusually rapid convergence, measured in orders of magnitude, does not signify superior performance, but rather is a typical overfitting phenomenon. In other words, the model lacks generalization ability and robustness, thus verifying the shortcomings of existing technologies in easily getting trapped in local optima and being unable to resist noise interference when physical constraints are lacking.

[0144] Table 1 presents quantitative comparison data of the model with introduced spectral norm constraints and the baseline model without introduced constraints provided by this invention in terms of three key indicators: mean square error, sparsity, and phase lag.

[0145] Table 1. Performance Comparison of Model Reconstruction Effects index Unconstrained Spectral norm constraints Increase Mean squared deviation 59.31 1.41 ↓40 times sparsity -0.0865 -2.0058 Successful separation Phase lag Fitting noise 0ms Real-time response As shown in Table 1, compared with the unconstrained baseline model, the method provided by this invention achieves significant improvements in all performance indicators. Specifically, in terms of the mean square error, this invention reduces the error value from 59.31 in the baseline model to 1.41, an improvement of approximately 40 times. This indicates that the spectral norm constraint effectively suppresses the interference of random noise on model fitting and significantly improves the accuracy of signal reconstruction. In terms of sparsity, the value changes from -0.0865 in the baseline model to -2.0058 in this invention, achieving a qualitative leap from inability to separate to successful separation. This verifies that this invention can accurately extract the residual components of high-frequency abrupt changes from the background signal. In terms of phase lag, this invention completely solves the phase disorder problem caused by fitting noise in the baseline model, achieving a real-time response of 0ms. This demonstrates that the hardware-software co-engineered architecture has extremely high sensitivity in capturing dynamic changes in the power grid and does not exhibit the lag phenomenon of traditional physical models.

[0146] Figure 11 This is a schematic diagram of the power time-series data prediction device provided by the present invention, as shown below. Figure 11 As shown, it mainly includes, but is not limited to: The first computing unit 1110 is used to acquire the physical baseline signal of power time series data.

[0147] Storage unit 1120, connected to the first computing unit 1110, is used to store the generation parameters of the physical baseline signal.

[0148] The second calculation unit 1130 is used to determine the residual signal of the power time series data based on the physical baseline signal of the power time series data; input the residual signal into the residual prediction model, and obtain the prediction residual signal output by the residual prediction model.

[0149] The physical isolation strip 1140 is used to block signal interaction between the first computing unit 1110 and the second computing unit 1130.

[0150] The gating unit 1150 is connected to the output of the first calculation unit 1110 and the output of the second calculation unit 1130. It is used to evaluate the prediction residual signal based on the prediction residual signal according to the validity determination mechanism, obtain the evaluation result, and decide, based on the evaluation result, to generate a prediction result for the power time series data based on the prediction residual signal and the physical baseline signal, or based only on the physical baseline signal.

[0151] It should be noted that the power time series data prediction device provided by the present invention can execute the power time series data prediction method described in any of the above embodiments during specific operation, and this embodiment will not elaborate on this.

[0152] The power time-series data prediction device provided by the present invention decouples power time-series data into physical baseline signals and residuals for separate processing, and uses an adaptive three-stage gating mechanism to effectively screen the prediction residual signals. This allows for the accurate capture of key signal spikes while preserving physical constraints to ensure power grid safety, effectively distinguishing between real load fluctuations and sensor noise, and avoiding the issuance of erroneous dispatch instructions due to misjudgment of noise.

[0153] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 12 As shown, the electronic device may include a processor 1210, a communications interface 1220, a memory 1230, and a communication bus 1240, wherein the processor 1210, the communications interface 1220, and the memory 1230 communicate with each other via the communication bus 1240. The processor 1210 can call logic instructions in the memory 1230 to execute a power time-series data prediction method, which includes: determining the residual signal of the power time-series data based on the physical baseline signal of the power time-series data; inputting the residual signal into a residual prediction model to obtain the predicted residual signal output by the residual prediction model; and, based on the magnitude of the predicted residual signal, deciding whether to generate a prediction result for the power time-series data based on the predicted residual signal and the physical baseline signal or only based on the physical baseline signal.

[0154] Furthermore, the logical instructions in the aforementioned memory 1230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the power time series data prediction method provided by the above methods. The method includes: determining the residual signal of the power time series data based on the physical baseline signal of the power time series data; inputting the residual signal into a residual prediction model to obtain the prediction residual signal output by the residual prediction model; and, based on the magnitude of the prediction residual signal, deciding to generate a prediction result for the power time series data based on the prediction residual signal and the physical baseline signal or based only on the physical baseline signal.

[0156] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program is implemented to perform the power time-series data prediction method provided by the above methods. The method includes: determining a residual signal of the power time-series data based on a physical baseline signal of the power time-series data; inputting the residual signal into a residual prediction model to obtain a predicted residual signal output by the residual prediction model; and, based on the magnitude of the predicted residual signal, deciding to generate a prediction result for the power time-series data based on the predicted residual signal and the physical baseline signal or based solely on the physical baseline signal.

[0157] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting power time-series data, characterized in that, include: Based on the physical baseline signal of the power time series data, determine the residual signal of the power time series data; The residual signal is input into the residual prediction model to obtain the predicted residual signal output by the residual prediction model; Based on the predicted residual signal, the predicted residual signal is evaluated according to the validity determination mechanism to obtain the evaluation result. Based on the evaluation result, a decision is made to generate a prediction result for the power time series data based on the predicted residual signal and the physical baseline signal, or based solely on the physical baseline signal.

2. The power time-series data prediction method according to claim 1, characterized in that, The physical baseline signal is determined based on the following steps: Construct the Hankel matrix of the power time-series data; Perform singular value decomposition on the Hankel matrix to obtain the left singular vector matrix, the central singular value matrix, and the right singular vector matrix; All singular values ​​of the central singular value matrix are sorted according to their numerical values ​​to obtain a singular value sequence, and a preset number of target singular values ​​are determined based on the singular value sequence. Based on the target singular value, the target left singular vector is determined from the left singular vector matrix, and the target right singular vector is determined from the right singular vector matrix; Based on the target singular values, the target left singular vector, and the target right singular vector, a low-rank matrix is ​​obtained; The low-rank matrix is ​​reconstructed to obtain the physical baseline signal.

3. The power time series data prediction method according to claim 2, characterized in that, The construction of the Hankel matrix for the power time-series data includes: The power time series data is scrolled and truncated based on a sliding window of preset length to obtain multiple continuous local time series subsequences; The Hankel matrix is ​​constructed based on the multiple consecutive local temporal subsequences, wherein the elements in the Hankel matrix remain consistent along the anti-diagonal direction.

4. The power time series data prediction method according to claim 1, characterized in that, The physical baseline signal is determined based on the following steps: Determine the wavelet basis functions and the target decomposition level; Based on the wavelet basis function, multi-scale discrete wavelet decomposition is performed on the power time series data to obtain an approximate set of wavelet coefficients located at the target decomposition level; The physical feature components are determined based on the set of approximate wavelet coefficients for the target decomposition level; The physical feature components are reconstructed by inverse wavelet transform to obtain the physical baseline signal.

5. The power time series data prediction method according to claim 1, characterized in that, The physical baseline signal is determined based on the following steps: Empirical mode decomposition is performed on the power time series data to obtain residual components; The physical baseline signal is determined based on the residual components.

6. The power time-series data prediction method according to claim 1, characterized in that, The step of inputting the residual signal into the residual prediction model and obtaining the predicted residual signal output by the residual prediction model includes: An initial weight matrix is ​​constructed based on the weights of all model parameters of the residual prediction model; Calculate the spectral norm of the initial weight matrix; The initial weight matrix is ​​normalized using the spectral norm to obtain a normalized weight matrix; The predicted residual signal is generated based on the normalized weight matrix and the residual signal.

7. The power time series data prediction method according to claim 1, characterized in that, The step of evaluating the predicted residual signal based on the prediction residual signal according to a validity determination mechanism to obtain an evaluation result, and then, based on the evaluation result, deciding whether to generate a prediction result for the power time series data based on the predicted residual signal and the physical baseline signal, or based solely on the physical baseline signal, includes: Based on the historical statistical characteristics within the sliding window of the power time-series data, a first noise tolerance threshold and a second safety saturation threshold are calculated, and the second safety saturation threshold is greater than the first noise tolerance threshold. Obtain the amplitude of the predicted residual signal, and perform the following decision logic based on the amplitude: If the amplitude is less than or equal to the first noise tolerance threshold, the predicted residual signal is determined to be invalid background noise, and the prediction result for the power time series data is generated only based on the physical baseline signal. If the amplitude is greater than the first noise tolerance threshold and less than the second safety saturation threshold, the predicted residual signal is determined to be a valid fluctuation signal, and the physical baseline signal and the predicted residual signal are weighted and fused based on the dynamic weighting coefficient to generate a prediction result for the power time series data; wherein, the dynamic weighting coefficient is determined based on the mapping position of the amplitude within the numerical range formed by the first noise tolerance threshold and the second safety saturation threshold; If the amplitude is greater than or equal to the second safety saturation threshold, the amplitude of the predicted residual signal is limited to obtain a safety residual signal, and a prediction result for the power time series data is generated based on the physical baseline signal and the safety residual signal.

8. The power time-series data prediction method according to claim 7, characterized in that, The step of weighting and fusing the physical baseline signal and the predicted residual signal based on dynamic weighting coefficients to generate a prediction result for the power time series data includes: Obtain the time-series index of the predicted residual signal; Based on the time-series index, the target baseline value of the physical baseline signal is determined; The predicted residual signal and the target baseline value are weighted and fused based on the dynamic weighting coefficients to obtain the prediction result of the power time series data.

9. The power time-series data prediction method according to claim 1, characterized in that, The training process of the residual prediction model includes: Obtain historical power time-series data; The historical power time series data is decomposed to obtain the historical physical baseline signal, and the historical residual sample is determined based on the difference between the historical power time series data and the historical physical baseline signal. Determine the residual labels corresponding to the historical residual samples; Freeze the generation parameters of the historical physical baseline signal and iteratively execute the training process until the preset termination condition is met; The training process includes: Input the historical residual samples into the residual prediction model to obtain the predicted residuals output by the residual prediction model; Calculate the loss function value between the predicted residual and the residual label; Based on the loss function value, the model parameters of the residual prediction model are updated using the backpropagation algorithm.

10. A power time-series data prediction device, characterized in that, include: The first computing unit is used to acquire the physical baseline signal of power time-series data; A storage unit, connected to the first computing unit, is used to store the generation parameters of the physical baseline signal; The second calculation unit is used to determine the residual signal of the power time series data based on the physical baseline signal of the power time series data; The residual signal is input into the residual prediction model to obtain the predicted residual signal output by the residual prediction model; A physical isolation zone is used to block signal interaction between the first computing unit and the second computing unit; A gating unit, connected to the output of the first computing unit and the output of the second computing unit, is used to evaluate the prediction residual signal based on the prediction residual signal according to the validity determination mechanism, obtain the evaluation result, and, based on the evaluation result, decide whether to generate a prediction result for the power time series data based on the prediction residual signal and the physical baseline signal, or based solely on the physical baseline signal.

11. The power time-series data prediction device according to claim 10, characterized in that, The power time-series data prediction device is implemented by a field-programmable gate array or an application-specific integrated circuit. The first computing unit and the second computing unit are respectively configured in different hardware logic partitions of the field-programmable gate array or the application-specific integrated circuit; The physical isolation strip is used to cut off the signal transmission path between the different hardware logical partitions.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the power time-series data prediction method as described in any one of claims 1 to 9.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the power time-series data prediction method as described in any one of claims 1 to 9.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the power time-series data prediction method as described in any one of claims 1 to 9.

Images