Electric quantity use condition prompting and safety early warning monitoring method and system
By using multi-source data fusion and joint optimization techniques from low-frequency metering channels and high-frequency harmonic channels, a structured state-space sequence model is constructed, which solves the accuracy and stability problems of power quality monitoring and prediction in existing technologies, and achieves high-precision prediction of power usage trends and safety early warning.
Patent Information
- Application Number
- CN202511351696.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-22
- Publication Date
- 2026-01-09
- Estimated Expiration
- 2045-09-22
AI Technical Summary
Existing power quality monitoring and prediction technologies have limitations when processing multi-channel fusion of high-frequency harmonics and low-frequency metering data. They cannot fully utilize time-series correlation characteristics, resulting in low prediction accuracy. Furthermore, they lack frequency domain consistency constraints and multi-task collaborative optimization, making it difficult to meet the needs of high-precision power quality management and safety early warning in complex operating environments.
By employing multi-source data fusion of low-frequency metering channels and high-frequency harmonic channels, parametric spectrum estimation, rotating harmonic state construction, gating fusion mechanism, and joint optimization technology for time-domain and frequency-domain consistency, a structured state-space sequence model is constructed to achieve harmonic feature extraction, model-driven prediction, and real-time early warning.
It improves the accuracy and stability of electricity usage trend prediction, enables high-precision power quality monitoring and safety early warning in complex power grid environments, and enhances the reliability and real-time performance of power grid operation.
Smart Images

Figure CN121308320A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of power quality monitoring and prediction, and particularly relates to a power usage situation prompting and safety early warning monitoring method and system. BACKGROUND
[0002] The existing power quality monitoring and prediction technology mainly relies on collecting and analyzing voltage and current waveforms in the process of power grid operation, extracting characteristic parameters such as harmonic frequency, amplitude and phase through signal processing methods such as Fourier transform, wavelet analysis and spectrum estimation, and evaluating power quality indicators such as total harmonic distortion and band-pass harmonic energy. In the prediction aspect, the traditional method mainly uses autoregressive moving average model, Kalman filter model and support vector regression model based on statistical modeling to fit and trend extrapolate the historical monitoring data. However, this kind of method has limitations in processing multi-channel fusion of high-frequency harmonics and low-frequency metering data, cannot fully utilize the time series correlation characteristics under different sampling frequencies, and has low prediction accuracy in the scene where the harmonic component changes significantly over time. In recent years, some researches introduce deep learning models such as recurrent neural network and time series convolution network to improve the non-linear modeling capability, but in the prediction of power quality driven by harmonics, there are still problems such as insufficient utilization of exogenous input features and lack of special optimization of model structure for harmonic time-varying characteristics.
[0003] In the harmonic parameter extraction link, the existing technology mainly regards harmonic analysis as a separate signal processing process, lacks deep coupling with the prediction model, and thus the harmonic characteristics cannot dynamically drive the prediction in the model hidden state updating process. In the prediction result optimization aspect, the existing method mainly relies on a single time domain error minimization objective, lacks frequency domain consistency constraints and multi-task collaborative optimization mechanism, and thus the prediction result deviates from the actual measurement in the frequency spectrum characteristics. For the prior range limitation of harmonic parameters and the feature-driven state space updating, the existing technology has not established an effective joint optimization framework, and the model has limited performance in dealing with harmonic feature changes, prediction uncertainty analysis and real-time monitoring and early warning, which leads to the fact that the existing harmonic state monitoring and multi-step prediction method is difficult to meet the demand of high-precision power quality management and safety early warning in terms of accuracy, stability and robustness in complex operating environment.
[0004] Therefore, how to provide a power usage situation prompting and safety early warning monitoring method and system is a problem to be solved by those skilled in the art. SUMMARY
[0005] One purpose of the present application is to propose an electric quantity use situation prompting and safety early warning monitoring method and system, the present application fully utilizes the multi-source data fusion of low-frequency metering channel and high-frequency harmonic channel, parameterized spectrum estimation, rotating harmonic state construction, gating fusion mechanism and time domain and frequency domain consistency joint optimization technology, details the complete process from harmonic feature extraction, model driven prediction to joint constraint optimization and real-time early warning, has the advantages of high prediction accuracy, strong spectrum feature retention, good adaptability to harmonic time-varying characteristics and high reliability of monitoring and early warning in complex power grid environment.
[0006] According to the electric quantity use situation prompting and safety early warning monitoring method of the embodiment of the present application, comprising:
[0007] Collecting low-frequency metering channel sampling data and high-frequency harmonic channel sampling data, preprocessing the high-frequency harmonic channel sampling data to obtain a high-frequency harmonic analysis window signal;
[0008] Performing parameterized spectrum estimation based on the high-frequency harmonic analysis window signal to obtain a harmonic parameter set containing harmonic order, harmonic frequency, harmonic amplitude and harmonic phase, and screening effective harmonic components;
[0009] Based on the harmonic parameter set, a rotating harmonic state sequence is constructed, adjacent analysis windows are time-smoothed, and total harmonic distortion and band-pass harmonic energy are calculated;
[0010] The low-frequency metering channel sampling data, total harmonic distortion and band-pass harmonic energy are taken as exogenous inputs to construct an improved structured state space sequence model, and future multi-step prediction results and model hidden states are output;
[0011] The rotating harmonic state sequence and the model hidden state are gate fused, time domain consistency constraints and frequency domain consistency constraints are imposed on the future multi-step prediction results, and harmonic frequency prior range and harmonic amplitude prior range are used to limit the search space of the harmonic parameter set;
[0012] According to the prediction error calculated based on the future multi-step prediction results and the actual measurement sequence, the time domain consistency constraints and the frequency domain consistency constraints, the rotating harmonic state sequence and the model hidden state, a joint optimization target is constructed, end-to-end training is performed, and the structured state space sequence model structure and the fusion layer parameters are solidified;
[0013] The structured state space sequence model is deployed to a monitoring platform to generate electric quantity use trend prediction results, harmonic index change trend and uncertainty parameters, and compared with preset safety threshold, when exceeding the limit, prompting and early warning are triggered and output to a visual terminal.
[0014] Optionally, the low-frequency metering channel sampling data specifically includes voltage effective value, current effective value, active power, reactive power and power factor, and the high-frequency harmonic channel sampling data specifically includes high-frequency sampled voltage waveform, current waveform and frequency spectrum component, wherein the low-frequency sampling frequency is 1 Hz to 5 Hz, and the high-frequency sampling frequency is 5 kHz to 20 kHz.
[0015] Optionally, the high-frequency harmonic channel sampling data is preprocessed to obtain a high-frequency harmonic analysis window signal, specifically referring to sequentially performing denoising, direct current component removal, band-pass filtering, segment interception according to a set window length and window processing on the high-frequency harmonic channel sampling data.
[0016] Optionally, parameterized spectrum estimation is performed based on the high-frequency harmonic analysis window signal to obtain a harmonic parameter set containing harmonic order, harmonic frequency, harmonic amplitude and harmonic phase, and valid harmonic components are screened, including:
[0017] The high-frequency harmonic analysis window signal is subjected to frequency domain decomposition to extract harmonic candidate components, and the harmonic order, harmonic frequency, harmonic amplitude and harmonic phase of each harmonic candidate component are calculated to form a first harmonic parameter set;
[0018] The trend change information in the low-frequency metering channel sampling data is subjected to cross-channel coupling operation with a high-frequency harmonic component set composed of harmonic components in the first harmonic parameter set, and the frequency and amplitude of the high-frequency harmonic component are offset corrected by the trend change information to obtain a second harmonic parameter set after correction;
[0019] While the second harmonic parameter set is generated, the change trend of the harmonic parameters in the adjacent historical analysis window is analyzed, and the harmonic parameter initial value of the current analysis window is predicted, the prediction result is taken as the initialization value of the second harmonic parameter set, the parameter search range is narrowed and the estimation convergence speed is accelerated;
[0020] The second harmonic parameter set after prediction initialization is used for iterative optimization to finely estimate the harmonic order, harmonic frequency, harmonic amplitude and harmonic phase of each harmonic component;
[0021] For each harmonic component in the second harmonic parameter set after iterative optimization, a comprehensive confidence index based on signal-to-noise ratio, spectral fitting residual and adjacent window time smoothing consistency is calculated;
[0022] The valid harmonic components meeting the confidence threshold condition are dynamically screened out, and the low-confidence or abnormal harmonic components are removed to obtain a third harmonic parameter set, and the third harmonic parameter set is taken as the final valid harmonic parameter set and output.
[0023] Optionally, the rotating harmonic state sequence is constructed based on the harmonic parameter set, time smoothing is performed on adjacent analysis windows, total harmonic distortion and band harmonic energy are calculated, and the method comprises the following steps:
[0024] Based on the harmonic parameter set, the harmonic order, harmonic frequency, harmonic amplitude and harmonic phase of each harmonic component are arranged in time sequence to generate a rotating harmonic state sequence corresponding to the analysis window, and the rotating harmonic state sequences of consecutive analysis windows are sequentially spliced to obtain a rotating harmonic state sequence set covering the target time range;
[0025] In the rotating harmonic state sequence set, time smoothing is performed on the rotating harmonic state sequences between adjacent analysis windows to generate a smoothed rotating harmonic state sequence;
[0026] Based on the smoothed rotating harmonic state sequence, the total harmonic distortion is calculated, and the total harmonic distortion is generated in a manner that the square root of the sum of the squares of the harmonic amplitudes is the numerator and the fundamental amplitude is the denominator;
[0027] Based on the smoothed rotating harmonic state sequence, the harmonic components are banded according to the harmonic order range to generate a harmonic amplitude sequence corresponding to each band, the band harmonic energy is calculated based on the harmonic amplitude sequence, and the band harmonic energy is generated by summing the squares of the harmonic amplitudes in the band. The total harmonic distortion and the band harmonic energy are recorded in time sequence.
[0028] Optionally, the low-frequency metering channel sampling data, the total harmonic distortion and the band harmonic energy are used as exogenous inputs to construct an improved structured state space sequence model, and the future multi-step prediction results and the model hidden state are outputted, comprising:
[0029] Based on the low-frequency metering channel sampling data, the total harmonic distortion and the band harmonic energy as exogenous inputs, an improved structured state space sequence model is constructed, which includes a harmonic state transition module, a harmonic gating decoding module and a multi-task output module;
[0030] In the harmonic state transition module, time-varying state transition parameters and exogenous input mapping parameters are generated based on the exogenous input vector and the initial hidden state of the model, and the model hidden state is updated, wherein the exogenous input vector is composed of the low-frequency metering channel sampling data, the total harmonic distortion and the band harmonic energy at the corresponding time step, and the initial hidden state of the model is calculated by the initial state mapping unit based on the feature vector of the low-frequency metering channel sampling data at the first analysis time step and the total harmonic distortion and the band harmonic energy at the corresponding time step;
[0031] In the harmonic gating decoding module, the updated model hidden state is connected with the total harmonic distortion and the banding harmonic energy feature vector, and then input into the gating function to generate a gating coefficient, and the updated model hidden state is gated and modulated based on the gating coefficient to obtain a gated and modulated hidden state representation;
[0032] In the multi-task output module, the hidden state representation and the exogenous input vector are input together for processing to generate future multi-step prediction results, and the prediction values corresponding to each prediction time step are output respectively, wherein each prediction time step corresponds to an independent readout parameter, and the prediction step number is determined by a preset prediction range;
[0033] The gating coefficient, the time-varying state transition parameter, and the exogenous input mapping parameter are recorded, and the future multi-step prediction results and the updated model hidden state are output and cached in time sequence.
[0034] Optionally, the gating fusion of the rotating harmonic state sequence and the model hidden state is performed to impose time domain consistency constraints and frequency domain consistency constraints on the future multi-step prediction results, and to generate a search space of harmonic frequency prior range and harmonic amplitude prior range to limit the harmonic parameter set, including:
[0035] The generated gating coefficient is called to perform element-wise weighted fusion on the model hidden state to form a gated and fused hidden state representation;
[0036] The current time, the corresponding harmonic amplitude, harmonic frequency, and harmonic phase parameters are extracted from the rotating harmonic state sequence, the reference current sequence is synthesized in the time domain, and the sum of the squared differences of the current prediction sequence and the reference current sequence in the future multiple time steps is calculated to obtain the time domain consistency constraint;
[0037] The current prediction sequence and the reference current sequence are respectively subjected to discrete Fourier transform to obtain a prediction spectrum and a reference spectrum; within a frequency neighborhood range centered on each target harmonic frequency and determined by a preset frequency neighborhood width, the sum of the absolute values of the differences between the prediction spectrum and the reference spectrum is calculated to obtain the frequency domain consistency constraint;
[0038] Based on the gated and fused hidden state representation, harmonic frequency prior values and harmonic amplitude prior values are generated, and a center value and a positive and negative range are set for each harmonic frequency prior value and harmonic amplitude prior value to form a frequency prior interval and an amplitude prior interval, thereby limiting the search space of the harmonic parameter set;
[0039] The gating coefficient, the gated and fused hidden state representation, the time domain consistency constraint, the frequency domain consistency constraint, and the frequency prior interval and the amplitude prior interval are recorded and saved together with the future multi-step prediction results.
[0040] The power usage situation prompting and safety early warning monitoring system according to the embodiment of the application comprises the following modules:
[0041] The acquisition and preprocessing module is used for acquiring low-frequency metering channel and high-frequency harmonic channel sampling data, and pre-processing the high-frequency harmonic channel sampling data to generate a high-frequency harmonic analysis window signal.
[0042] The harmonic parameter module is used for performing parameterized spectrum estimation based on the high-frequency harmonic analysis window signal to screen effective harmonic components.
[0043] The state sequence construction module is used for generating a rotating harmonic state sequence based on the harmonic parameter set, performing time smoothing on adjacent analysis windows, and calculating total harmonic distortion and banding harmonic energy.
[0044] The state space modeling module is used for constructing an improved structured state space sequence model to output future multi-step prediction results and model hidden states.
[0045] The fusion constraint module is used for fusing the rotating harmonic state sequence and the model hidden state, and imposing time domain consistency and frequency domain consistency constraints on the future multi-step prediction results.
[0046] The joint optimization module is used for calculating prediction errors according to the future multi-step prediction results and actual measurement sequences, constructing a joint optimization target, and solidifying the structured state space sequence model structure and the fusion layer parameters.
[0047] The monitoring and early warning module is used for deploying the structured state space sequence model to a monitoring platform, triggering a prompt and an early warning, and outputting to a visual terminal.
[0048] The application has the following advantages:
[0049] The application introduces a structured state space sequence model and a rotating harmonic state driven multi-source fusion method, realizes collaborative modeling of low-frequency metering data and high-frequency harmonic characteristics, captures power quality change trends, and accurately depicts the dynamic evolution process of harmonics.
[0050] Through a joint optimization strategy, prediction errors, time domain consistency constraints, and frequency domain consistency constraints are included in a unified optimization framework, so that the model can simultaneously consider the fitting effects of the time domain and the frequency domain during the training process, and the adaptability to power grid harmonic time-varying characteristics is improved.
[0051] The application can not only provide high-precision power usage trend and harmonic index change prediction, but also timely trigger early warning when the prediction result exceeds the safety threshold, thereby enhancing the real-time performance and reliability of power quality monitoring and providing strong guarantee for the safety and stability of power grid operation. BRIEF DESCRIPTION OF DRAWINGS
[0052] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are used to explain the application, but do not constitute a limitation on the application. In the drawings:
[0053] Figure 1 A flowchart of an electric quantity usage situation prompting and safety early warning monitoring method is provided for the application;
[0054] Figure 2 A structural schematic diagram of an electric quantity usage situation prompting and safety early warning monitoring system is provided for the application. DETAILED DESCRIPTION
[0055] The application will now be described in further detail with reference to the drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.
[0056] REFERENCE Figure 1 An electric quantity usage situation prompting and safety early warning monitoring method comprises:
[0057] Collect low-frequency metering channel sampling data and high-frequency harmonic channel sampling data, pre-process the high-frequency harmonic channel sampling data to obtain a high-frequency harmonic analysis window signal;
[0058] Perform parameterized spectrum estimation based on the high-frequency harmonic analysis window signal to obtain a harmonic parameter set containing harmonic order, harmonic frequency, harmonic amplitude and harmonic phase, and filter effective harmonic components;
[0059] Construct a rotating harmonic state sequence based on the harmonic parameter set, time-smooth adjacent analysis windows, and calculate total harmonic distortion and banded harmonic energy;
[0060] Take the low-frequency metering channel sampling data, total harmonic distortion and banded harmonic energy as exogenous inputs, construct an improved structured state space sequence model, and output future multi-step prediction results and model hidden states;
[0061] Gate fusion is performed on the rotating harmonic state sequence and the model hidden state, time-domain consistency constraints and frequency-domain consistency constraints are imposed on the future multi-step prediction results, and harmonic frequency prior range and harmonic amplitude prior range are generated to limit the search space of the harmonic parameter set;
[0062] The prediction error is calculated according to the future multi-step prediction result and the actual measurement sequence, combined with the time domain consistency constraint and the frequency domain consistency constraint, and the rotating harmonic state sequence and the model hidden state, a joint optimization target is constructed, end-to-end training is performed, and the structured state space sequence model structure and the fusion layer parameters are solidified;
[0063] The structured state space sequence model is deployed to the monitoring platform to generate power usage trend prediction results, harmonic index change trend and uncertainty parameters, and compared with the preset safety threshold, and when the limit is exceeded, a prompt and warning are triggered and output to the visualization terminal.
[0064] In the embodiment, the low-frequency metering channel sampling data specifically includes voltage effective value, current effective value, active power, reactive power and power factor, and the high-frequency harmonic channel sampling data specifically includes high-frequency sampled voltage waveform, current waveform and frequency spectrum component, wherein the low-frequency sampling frequency is 1Hz to 5Hz, and the high-frequency sampling frequency is 5kHz to 20kHz.
[0065] In the embodiment, the high-frequency harmonic channel sampling data is preprocessed to obtain a high-frequency harmonic analysis window signal, specifically referring to sequentially performing denoising, removing direct current component, band-pass filtering, and segmenting and intercepting according to the set window length and window processing on the high-frequency harmonic channel sampling data.
[0066] In the embodiment, the parameterized spectrum estimation is performed based on the high-frequency harmonic analysis window signal to obtain a harmonic parameter set containing harmonic order, harmonic frequency, harmonic amplitude and harmonic phase, and effective harmonic components are screened, including:
[0067] The high-frequency harmonic analysis window signal is frequency domain decomposed to extract harmonic candidate components, and the harmonic order, harmonic frequency, harmonic amplitude and harmonic phase of each harmonic candidate component are calculated respectively to form a first harmonic parameter set, wherein the harmonic order is obtained by rounding the ratio of the harmonic frequency to the fundamental frequency, the harmonic frequency is obtained by multiplying the harmonic order by the fundamental frequency, the harmonic amplitude is obtained by taking the modulus of the frequency domain amplitude-frequency characteristic of the harmonic candidate component, and the harmonic phase is obtained by taking the inverse tangent of the frequency domain phase-frequency characteristic of the harmonic candidate component;
[0068] The trend change information in the low-frequency metering channel sampling data and the high-frequency harmonic component set composed of the harmonic components in the first harmonic parameter set are cross-channel coupled and operated, the frequency and amplitude of the high-frequency harmonic component are offset corrected by the trend change information, and a second harmonic parameter set is obtained.
[0069] While generating the second set of harmonic parameters, the changing trends of harmonic parameters within adjacent historical analysis windows are analyzed, and the initial values of harmonic parameters in the current analysis window are predicted. The prediction results are used as the initial values of the second set of harmonic parameters, narrowing the parameter search range and accelerating the estimation convergence speed. Specifically:
[0070] When generating the second harmonic parameter set, the frequency, amplitude, and phase estimation results of each harmonic component within adjacent historical analysis windows are first extracted, the changing trends are calculated, and the initial values of the harmonic parameters in the current analysis window are predicted using linear extrapolation. The frequency and amplitude are predicted using the differential trends of the previous two windows, while the phase is calculated based on the phase and frequency of the previous window and processed modulo 2π. The prediction results are used as the initial values for the second harmonic parameter set, effectively narrowing the parameter search range and reducing invalid calculations.
[0071] Frequency prediction:
[0072] ;
[0073] Amplitude prediction:
[0074] ;
[0075] Phase prediction:
[0076] ;
[0077] in, , , These represent the current analysis window. Inner Initialization results of the frequency, amplitude, and phase of each harmonic component. , These are the frequency estimates for the previous window and the two previous windows, respectively. , These are the amplitude estimates for the previous window and the two windows prior, respectively. This is the phase estimate from the previous window. This is the length of the high-frequency analysis window. For high frequency sampling rate, , These are the trend weighting coefficients for frequency and amplitude prediction, respectively. Operations are used to maintain phase in Continuity of intervals;
[0078] By extracting the estimated frequency, amplitude, and phase of each harmonic component within adjacent historical analysis windows and calculating their changing trends, the algorithm predicts the initial values of harmonic parameters for the current analysis window using linear extrapolation. This improves initialization accuracy and computational efficiency. Frequency and amplitude predictions rely on the changing trends of the previous two windows, effectively capturing dynamic changes in harmonic characteristics. Phase prediction combines the phase and frequency calculation results from the previous window and maintains phase continuity within a reasonable range through modulo operations, avoiding estimation errors caused by phase jumps. This reduces the deviation between initial and true values, narrows the range of subsequent parameter searches, and reduces unnecessary computation, thereby accelerating the convergence speed of the estimation process. By introducing a trend weight coefficient to adjust the prediction sensitivity, the algorithm maintains high stability and accuracy under different dynamic conditions, improving the real-time performance and reliability of harmonic analysis and prediction.
[0079] Iterative optimization is performed using the predicted and initialized set of second harmonic parameters to finely estimate the harmonic order, harmonic frequency, harmonic amplitude, and harmonic phase of each harmonic component. Specifically:
[0080] The harmonic order is fine-tuned based on the difference between the predicted and measured values, and the order correction is achieved by minimizing the frequency domain error function.
[0081] High-resolution spectral analysis methods are used to refine the calculation of harmonic frequencies, thereby improving the accuracy of frequency estimation.
[0082] By combining weighted least squares fitting, harmonic amplitude and harmonic phase are jointly optimized to reduce amplitude and phase estimation errors.
[0083] For each harmonic component in the iteratively optimized second harmonic parameter set, a comprehensive confidence index based on signal-to-noise ratio, spectral fitting residual, and consistency of temporal smoothing between adjacent windows is calculated. The comprehensive confidence index is: Among them, the comprehensive confidence index serves as a scoring standard for determining the validity of harmonic components, directly determining whether a harmonic component enters the third harmonic parameter set:
[0084] ;
[0085] in, Indicates the first Signal-to-noise ratio of each harmonic component Indicates the first The spectral fitting residuals of each harmonic component Indicates the first A measure of the inconsistency in time smoothing between each harmonic component and adjacent analysis windows. These are the weighting coefficients. , , respectively, are the maximum values of the corresponding indicators in all harmonic components, used for normalization processing;
[0086] The effective harmonic components satisfying the confidence threshold condition are dynamically screened out, the low-confidence or abnormal harmonic components are removed, a third harmonic parameter set is obtained, and the third harmonic parameter set is output as the final effective harmonic parameter set.
[0087] The present application realizes high-precision and high-efficiency estimation of the harmonic parameter set by introducing the double mechanism of cross-channel coupling correction and trend prediction initialization. When generating the second harmonic parameter set, the trend information of the low-frequency measurement channel is used to offset and correct the frequency and amplitude of the high-frequency harmonic component, effectively suppressing the parameter estimation deviation caused by measurement noise, transient fluctuation or sampling drift, and improving the parameter stability and anti-interference ability. The parameter change trend of the adjacent historical analysis window is used for linear extrapolation prediction, and the prediction result is used as the initialization value, which reduces the parameter search range, reduces the invalid iteration calculation, accelerates the convergence speed of the iteration optimization process and reduces the calculation load. In the parameter optimization stage, high-resolution spectral analysis and weighted least squares fitting are combined to improve the estimation accuracy of the harmonic frequency, amplitude and phase. The comprehensive confidence index based on signal-to-noise ratio, spectral fitting residual and time smoothing consistency is introduced to dynamically screen the harmonic components, which can remove abnormal or low-confidence components while ensuring the retention of effective information, thereby improving the accuracy and reliability of the final harmonic parameter set.
[0088] In the embodiment, based on the harmonic parameter set, a rotating harmonic state sequence is constructed, adjacent analysis windows are time smoothed, total harmonic distortion and sub-band harmonic energy are calculated, including:
[0089] Based on the harmonic parameter set, the harmonic order, harmonic frequency, harmonic amplitude and harmonic phase of each harmonic component are arranged in time sequence to generate a rotating harmonic state sequence corresponding to the analysis window, and the rotating harmonic state sequences of consecutive analysis windows are sequentially spliced to obtain a rotating harmonic state sequence set covering the target time range.
[0090] In the rotating harmonic state sequence set, time smoothing processing is performed on the rotating harmonic state sequences between adjacent analysis windows to generate a smoothed rotating harmonic state sequence.
[0091] Based on the smoothed rotating harmonic state sequence, the total harmonic distortion is calculated, and the total harmonic distortion is generated in a manner that the square root of the sum of the squares of the harmonic amplitudes is the numerator and the fundamental amplitude is the denominator.
[0092] Based on the smoothed rotating harmonic state sequence, the harmonic components are banded according to the harmonic order range to generate the corresponding harmonic amplitude sequence, the banded harmonic energy is calculated based on the harmonic amplitude sequence, and the banded harmonic energy is generated by summing the squares of the harmonic amplitudes in the band, the total harmonic distortion and the banded harmonic energy are recorded according to the time sequence, and the banded harmonic energy is calculated based on the harmonic amplitude sequence. Specifically, the squares of the amplitudes of all harmonic components in the same harmonic order range are summed to reflect the total energy of the frequency band.
[0093] The application realizes continuous and stable expression of harmonic characteristics by constructing a rotating harmonic state sequence and performing time smoothing. In the harmonic state construction stage, the harmonic order, frequency, amplitude and phase are organized in time sequence as a rotating sequence, and the continuous analysis window is sequentially spliced, which can effectively retain the time sequence evolution characteristics of the harmonic parameters and avoid the one-sidedness of single window analysis. By performing time smoothing on the adjacent window state sequence, the impact of instantaneous disturbance and random noise on the harmonic characteristics can be suppressed, making the harmonic change trend more smooth and real. Based on the smoothed state sequence, the total harmonic distortion is calculated, and the fundamental amplitude normalization processing is adopted to ensure the comparability and stability under different operating conditions. The harmonics are banded according to the order range, and the banded harmonic energy is calculated, which can accurately analyze the influence of harmonics in different frequency bands on the system, help to locate the high-risk frequency band and carry out targeted governance, retain the dynamic characteristics of harmonics, and improve the stability, explainability and analysis accuracy of harmonic indicators. It provides a reliable quantitative basis for power quality monitoring, fault diagnosis and predictive maintenance, and has strong engineering practical value and popularization potential.
[0094] In the embodiment, the low-frequency metering channel sampling data, the total harmonic distortion and the banded harmonic energy are used as exogenous inputs to construct an improved structured state space sequence model, and the future multi-step prediction results and the model hidden state are output, including:
[0095] Based on the low-frequency metering channel sampling data, the total harmonic distortion and the banded harmonic energy as exogenous inputs, an improved structured state space sequence model is constructed, including a harmonic state transition module, a harmonic gating decoding module and a multi-task output module.
[0096] In the harmonic state transition module, time-varying state transition parameters and exogenous input mapping parameters are generated with the exogenous input vector and the initial hidden state of the model as variables, and the model hidden state is updated, wherein the exogenous input vector is composed of the low-frequency metering channel sampling data, the total harmonic distortion and the banded harmonic energy at the corresponding time step, the initial hidden state of the model is calculated by the feature vector of the low-frequency metering channel sampling data at the first analysis time step and the total harmonic distortion and the banded harmonic energy at the corresponding time step through the initial state mapping unit, and the model hidden state is updated.
[0097] inputting the exogenous input vector and the current model hidden state into a state transition function to generate time-varying state transition parameters and exogenous input mapping parameters corresponding to the time step;
[0098] transforming the hidden state of the previous time step according to the time-varying state transition parameters, mapping the current exogenous input vector using the exogenous input mapping parameters, and superimposing the results of the two to obtain an intermediate update value of the hidden state;
[0099] sequentially performing nonlinear activation processing and normalization processing on the intermediate update value of the hidden state to obtain an updated model hidden state;
[0100] In the harmonic gate decoding module, the updated model hidden state and the total harmonic distortion and banding harmonic energy feature vector are connected and input into a gating function to generate a gating coefficient, and the updated model hidden state is modulated based on the gating coefficient to obtain a gated hidden state representation, wherein the gating function is a function that outputs a gating coefficient with a value range of 0 to 1 by performing weighted summation on the input feature vector and applying nonlinear activation operation;
[0101] In the multi-task output module, the hidden state representation and the exogenous input vector are input together to generate future multi-step prediction results, and the prediction values corresponding to each prediction time step are output respectively, wherein each prediction time step corresponds to an independent readout parameter, and the prediction step number is determined by a preset prediction range. The hidden state representation and the exogenous input vector are input together, specifically as follows:
[0102] The hidden state representation and the exogenous input vector are spliced according to the feature dimension to form a joint feature representation;
[0103] The joint feature representation is sequentially input into a fully connected mapping layer and a nonlinear activation layer to extract deep features required for prediction;
[0104] The extracted deep features are input into the independent readout layer corresponding to each prediction time step respectively to output the prediction values of each time step;
[0105] The gating coefficient, the time-varying state transition parameters and the exogenous input mapping parameters are recorded, and the future multi-step prediction results and the updated model hidden state are output and cached in time sequence.
[0106] The application fuses low-frequency metering channel sampling data and total harmonic distortion, sub-band harmonic energy into exogenous input, constructs an improved structured state space sequence model with time-varying state transition ability, gating modulation ability and multi-task output ability, and realizes joint modeling of harmonic state change and power use trend. The harmonic state transition module can dynamically adjust the state transition parameters according to the exogenous input, so that the model has self-adaptability to different operating conditions and harmonic characteristics; the harmonic gating decoding module can adjust the hidden state contribution according to the importance of harmonic characteristics, and effectively suppresses irrelevant or noise information interference; the multi-task output module improves the accuracy and stability of short-term and long-term prediction by jointly predicting the results of multiple time steps. The records of gating coefficients, state transition parameters and exogenous input mapping parameters provide explainability support for subsequent analysis, improve the description accuracy of harmonic time sequence characteristics and the adaptability to non-stationary signals of the prediction model, and realize high-precision, multi-step and interpretable harmonic trend prediction.
[0107] In the embodiment, the rotating type harmonic state sequence and the model hidden state are gated and fused, time domain consistency constraints and frequency domain consistency constraints are imposed on the future multi-step prediction results, and the search space of the harmonic parameter set is generated by generating the harmonic frequency prior range and the harmonic amplitude prior range, including:
[0108] The generated gating coefficients are called to perform element-wise weighted fusion on the model hidden state to form a gated fusion hidden state representation;
[0109] The current time is extracted from the future multi-step prediction results, the corresponding harmonic amplitude, harmonic frequency and harmonic phase parameters are extracted from the rotating type harmonic state sequence at the current time, the reference current sequence is synthesized in the time domain, and the sum of the squared differences of the current prediction sequence and the reference current sequence at the future multiple time steps is calculated to obtain the time domain consistency constraint;
[0110] The current prediction sequence and the reference current sequence are respectively subjected to discrete Fourier transform to obtain the prediction spectrum and the reference spectrum; in the frequency neighborhood range centered on each target harmonic frequency and determined by the preset frequency neighborhood width, the sum of the absolute values of the differences between the prediction spectrum and the reference spectrum is calculated to obtain the frequency domain consistency constraint, wherein the frequency neighborhood width is a preset positive numerical parameter;
[0111] Based on the gated fusion hidden state representation, the harmonic frequency prior value and the harmonic amplitude prior value are generated, and a center value and a positive and negative range are set for each harmonic frequency prior value and harmonic amplitude prior value to form a frequency prior interval and an amplitude prior interval, thereby limiting the search space of the harmonic parameter set;
[0112] The gating coefficient, the gating fusion hidden state representation, the time domain consistency constraint, the frequency domain consistency constraint, and the frequency prior interval and the amplitude prior interval are recorded and saved together with the future multi-step prediction result.
[0113] The present application realizes multi-dimensional accurate constraint and search space optimization of the prediction result by introducing time domain and frequency domain double consistency constraints and combining the harmonic parameter prior interval generated by the gating fusion hidden state. The gating fusion hidden state representation can fully utilize the gating coefficient to weight and strengthen useful information and suppress irrelevant features at the feature level, improving the expression ability of the hidden state to the harmonic dynamic characteristics. The time domain consistency constraint ensures that the current prediction sequence is highly matched with the reference waveform synthesized by the harmonic parameters in the time waveform, which helps to improve the waveform fidelity of the prediction. The frequency domain consistency constraint matches the prediction and the reference in the local frequency neighborhood in the frequency spectrum structure, which can effectively suppress the frequency domain drift and the deviation of the harmonic energy distribution. The harmonic frequency and amplitude prior interval generated by the hidden state limits the search range, which not only reduces the solution space of the subsequent parameter estimation and reduces invalid calculation, but also significantly speeds up the convergence speed and improves the stability of global optimization. The present application guarantees the prediction accuracy while considering the calculation efficiency and model robustness, enhancing the reliability and interpretability of harmonic analysis and prediction.
[0114] In the present embodiment, the prediction error is calculated according to the multi-step prediction result and the actual measurement sequence, the joint optimization target is constructed by combining the time domain consistency constraint and the frequency domain consistency constraint, the rotating harmonic state sequence and the model hidden state, the end-to-end training is performed, and the structured state space sequence model structure and the fusion layer parameters are solidified, including:
[0115] The prediction error is calculated according to the multi-step prediction result and the actual measurement sequence, the joint optimization target function is constructed by combining the time domain consistency constraint, the frequency domain consistency constraint, the rotating harmonic state sequence constraint and the model hidden state constraint :
[0116] ;
[0117] Wherein, is a weight coefficient, is a prediction error, is a time domain consistency constraint, is a frequency domain consistency constraint, is a comprehensive measure of the rotating harmonic state sequence constraint and the model hidden state constraint;
[0118] The gradient-based optimization algorithm is used to iteratively solve the joint optimization target function, and the structured state space sequence model parameters are updated in each iteration;
[0119] When the joint optimization process reaches the convergence condition or reaches the preset upper limit of the number of iterations, output the structured state space sequence model parameters obtained by final optimization.
[0120] In this embodiment, the structured state space sequence model is deployed to the monitoring platform to generate power usage trend prediction results, harmonic index change trends, and uncertainty parameters, and is compared with preset safety thresholds. When the thresholds are exceeded, a prompt and warning are triggered and output to a visualization terminal, including:
[0121] The solidified structured state space sequence model is deployed to the monitoring platform to receive low-frequency metering channel sampling data and high-frequency harmonic channel sampling data transmitted in real time by the collection preprocessing module, and generate power usage trend prediction result sequences for corresponding time periods , wherein, is the predicted step active power value;
[0122] The harmonic index change trend is calculated based on the power usage trend prediction result sequence, wherein the total harmonic distortion rate is :
[0123] ;
[0124] , wherein, is the predicted step harmonic voltage effective value, is the fundamental voltage effective value, is the upper limit of the harmonic order;
[0125] The uncertainty parameter in the prediction result is calculated :
[0126] ;
[0127] , wherein, is the step sampling prediction value, is the sample mean of M times of sampling prediction values, is the number of samplings;
[0128] The power usage trend prediction results, harmonic index change trends, and uncertainty parameters are compared with preset safety thresholds , , respectively. When there is a prediction step or or , a prompt and warning mechanism are triggered;
[0129] The triggered prompt and early warning information are output through a visual terminal, including a prediction time period, a prediction value, a threshold value, an overrun type and a level, and corresponding data is recorded in a monitoring platform database.
[0130] Reference Figure 2 A power usage situation prompt and safety early warning monitoring system includes the following modules:
[0131] A collection preprocessing module is configured to collect low-frequency metering channel and high-frequency harmonic channel sampling data, and to preprocess the high-frequency harmonic channel sampling data to generate a high-frequency harmonic analysis window signal.
[0132] A harmonic parameter module is configured to perform parameterized spectrum estimation based on the high-frequency harmonic analysis window signal to screen effective harmonic components.
[0133] A state sequence construction module is configured to generate a rotating harmonic state sequence based on a harmonic parameter set, to perform time smoothing on adjacent analysis windows, and to calculate total harmonic distortion and band harmonic energy.
[0134] A state space modeling module is configured to construct an improved structured state space sequence model to output future multi-step prediction results and model hidden states.
[0135] A fusion constraint module is configured to gate fuse the rotating harmonic state sequence and the model hidden state, and to impose time domain consistency and frequency domain consistency constraints on the future multi-step prediction results.
[0136] A joint optimization module is configured to calculate prediction errors based on the future multi-step prediction results and actual measurement sequences, to construct a joint optimization target, and to solidify the structured state space sequence model structure and the fusion layer parameters.
[0137] A monitoring and early warning module is configured to deploy the structured state space sequence model to a monitoring platform, to trigger a prompt and an early warning, and to output to a visual terminal.
[0138] Embodiment 1
[0139] To verify the feasibility of the application in implementation, the application is applied to a certain industrial park A power distribution room, the typical load of the park is composed of multiple numerical control machine tools, injection molding machines and two high-power variable frequency calendering equipment, the peak load fluctuates obviously in the morning and evening of weekdays, and the start and stop of high-frequency electrical equipment leads to the lifting of harmonic level. Previously, the traditional single-channel time sequence prediction method used in the park can only extrapolate the load trend based on low-frequency metering data, and cannot timely perceive the rapid change of harmonic components, often resulting in large prediction deviation of total harmonic distortion (THD) and lagging or even missing of early warning, especially during the production switching period in the afternoon of Wednesday and Friday, harmonic out-of-limit events are more concentrated. In view of this pain point, the application is based on low-frequency metering channel sampling data and high-frequency harmonic channel sampling data, and constructs a whole-link process from harmonic parameterized spectrum estimation to rotating harmonic state sequence, and then to improved structured state space sequence model, trying to accurately capture the time-varying characteristics of harmonics and give early risk warning while ensuring the accuracy of power trend prediction.
[0140] In field deployment, the low-frequency metering channel samples the voltage effective value, current effective value, active power, reactive power and power factor at 1 Hz, and the high-frequency harmonic channel samples the voltage and current raw waveform at 5120 Hz, and after denoising, removing direct current, bandpass and windowing preprocessing, the high-frequency harmonic analysis window signal is formed. The parameterized spectrum estimation process outputs a harmonic parameter set containing harmonic order, harmonic frequency, harmonic amplitude and harmonic phase; to avoid noise interference, components with amplitude less than 0.05% of the fundamental wave amplitude are removed. Smooth adjacent windows in time to get a rotating harmonic state sequence, and calculate the total harmonic distortion and band harmonic energy. These two types of harmonic indicators and low-frequency metering channel sampling data are used as exogenous input to jointly drive the improved structured state space sequence model; the model integrates the rotating harmonic state sequence and the model hidden state in the gating fusion, and simultaneously generates the prediction values of the power trend and harmonic indicators in the next 30 minutes on the output side, and gives the uncertainty parameter for threshold comparison and graded early warning.
[0141] To verify the beneficial effects, the historical data from March to May 2025 is used for training, and the data in June is used for online verification. The comparison criteria are selected as: the traditional LSTM baseline model under the same site, same period and same sampling configuration. The evaluation indexes include THD prediction error (absolute error, unit %), power trend prediction error (MAPE, unit %) and the advance and accuracy of early warning triggering.
[0142] Table 1: Monitoring and prediction sample summary of industrial park A site (June 2025)
[0143]
[0144] From the data in Table 1, it can be seen that the method is stable in predicting the total harmonic distortion rate (THD) and power under different dates and time periods. For example, on June 5, 2025, from 08:00 to 08:30, the actual THD is 4.6%, and the predicted value is 4.5%. The THD error of the method is only 0.1%, the power MAPE (mean absolute percentage error) is 0.6%, and no warning is triggered. Similarly, under multiple working conditions such as June 7 and June 15, the THD error is controlled within 0.1%, and the power MAPE is less than 0.6%, indicating that the model has high prediction accuracy and stability within the range of normal load.
[0145] On June 12, from 14:00 to 14:30, and on June 18, from 15:30 to 16:00, the actual THD reached 5.1% and 6.3% respectively, both exceeding the safety threshold, so the system triggered a warning. However, the THD error of the method under high load conditions still remained between 0.1% and 0.2%, and the power MAPE was controlled within 1.0%, showing strong prediction robustness under extreme conditions.
[0146] Overall, among the 8 groups of data, 6 groups were in a non-warning state, and the prediction error remained at a very low level. Among the 2 groups that triggered a warning, the deviation between the predicted results and the actual values was also very small, fully demonstrating that the method can still maintain high-precision prediction in uncertain environments and can trigger harmonic anomaly alarms in time, which is helpful for early intervention in power safety, and is not only suitable for daily power quality monitoring, but also has practical value in peak load and harmonic anomaly scenarios.
[0147] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.
Claims
1. A method for monitoring and indicating power usage status and providing safety warnings, characterized in that, include: Collect sampling data from the low-frequency metering channel and the high-frequency harmonic channel. Preprocess the high-frequency harmonic channel sampling data to obtain the high-frequency harmonic analysis window signal. Based on the high-frequency harmonic analysis window signal, parameterized spectrum estimation is performed to obtain a set of harmonic parameters including harmonic order, harmonic frequency, harmonic amplitude, and harmonic phase, and effective harmonic components are screened. A rotating harmonic state sequence is constructed based on the harmonic parameter set. Time smoothing is performed on adjacent analysis windows to calculate total harmonic distortion and zonal harmonic energy. Using low-frequency metering channel sampling data, total harmonic distortion, and zonal harmonic energy as exogenous inputs, an improved structured state-space sequence model is constructed, and the future multi-step prediction results and model hidden states are output. Gated fusion of rotating harmonic state sequences and model latent states is performed, time-domain consistency constraints and frequency-domain consistency constraints are applied to future multi-step prediction results, and a search space is generated that limits the harmonic parameter set by the prior range of harmonic frequency and the prior range of harmonic amplitude. Based on the prediction results of future multi-step predictions and the actual measurement sequence, the prediction error is calculated. Combining the time domain consistency constraints and frequency domain consistency constraints, as well as the rotating harmonic state sequence and the model hidden state, a joint optimization objective is constructed. End-to-end training is performed and the structured state space sequence model structure and fusion layer parameters are solidified. The structured state-space sequence model is deployed to the monitoring platform to generate power consumption trend prediction results, harmonic index change trends and uncertainty parameters, and compare them with preset safety thresholds. When the limit is exceeded, prompts and warnings are triggered and output to the visualization terminal.
2. The method for monitoring and alerting power usage as claimed in claim 1, characterized in that, The low-frequency metering channel sampling data specifically includes the effective value of voltage, the effective value of current, active power, reactive power, and power factor. The high-frequency harmonic channel sampling data specifically includes the voltage waveform, current waveform, and spectral components sampled at high frequencies. The low-frequency sampling frequency is 1Hz to 5Hz, and the high-frequency sampling frequency is 5kHz to 20kHz.
3. The method for monitoring and alerting power usage status and providing safety warnings according to claim 1, characterized in that, The preprocessing of the high-frequency harmonic channel sampling data to obtain the high-frequency harmonic analysis window signal specifically refers to sequentially performing noise reduction, DC component removal, bandpass filtering, and segmentation and windowing processing on the high-frequency harmonic channel sampling data according to the set window length.
4. The method for monitoring and alerting power usage status and providing safety warnings according to claim 1, characterized in that, The parameterized spectrum estimation based on the high-frequency harmonic analysis window signal is performed to obtain a set of harmonic parameters including harmonic order, harmonic frequency, harmonic amplitude, and harmonic phase, and to filter effective harmonic components, including: The high-frequency harmonic analysis window signal is decomposed in the frequency domain to extract the candidate harmonic components. The harmonic order, harmonic frequency, harmonic amplitude and harmonic phase of each candidate harmonic component are calculated to form the first set of harmonic parameters. The trend change information in the low-frequency metering channel sampling data is coupled across channels with the high-frequency harmonic component set composed of harmonic components in the first harmonic parameter set. The frequency and amplitude of the high-frequency harmonic components are offset and corrected by the trend change information to obtain the corrected second harmonic parameter set. While generating the second set of harmonic parameters, the changing trend of harmonic parameters in adjacent historical analysis windows is analyzed, and the initial values of harmonic parameters in the current analysis window are predicted. The prediction results are used as the initial values of the second set of harmonic parameters, which narrows the parameter search range and speeds up the estimation convergence. Iterative optimization is performed using the predicted and initialized set of second harmonic parameters to make fine estimates of the harmonic order, harmonic frequency, harmonic amplitude, and harmonic phase of each harmonic component. For each harmonic component in the iteratively optimized second harmonic parameter set, a comprehensive confidence index based on signal-to-noise ratio, spectral fitting residual, and consistency of temporal smoothing between adjacent windows is calculated. The effective harmonic components that meet the confidence threshold are dynamically selected, and low-confidence or abnormal harmonic components are removed to obtain the third harmonic parameter set. The third harmonic parameter set is then output as the final effective harmonic parameter set.
5. The method for monitoring and alerting power usage as claimed in claim 1, characterized in that, The process involves constructing a rotating harmonic state sequence based on a set of harmonic parameters, performing time smoothing on adjacent analysis windows, and calculating total harmonic distortion and zonal harmonic energy, including: Based on the set of harmonic parameters, the harmonic order, harmonic frequency, harmonic amplitude and harmonic phase of each harmonic component are arranged in chronological order to generate a rotating harmonic state sequence corresponding to the analysis window. The rotating harmonic state sequences are then sequentially spliced to obtain a set of rotating harmonic state sequences covering the target time range. In the set of rotating harmonic state sequences, time smoothing is performed on the rotating harmonic state sequences between adjacent analysis windows to generate smoothed rotating harmonic state sequences; Total harmonic distortion (THD) is calculated based on a smoothed rotating harmonic state sequence. The THD is generated by using the fundamental amplitude as the denominator and the square root of the sum of the squares of the amplitudes of each harmonic as the numerator. Based on the smoothed rotating harmonic state sequence, the harmonic components are divided into zones according to the harmonic order range, generating a harmonic amplitude sequence corresponding to each zone. The zone harmonic energy is calculated based on the harmonic amplitude sequence. The zone harmonic energy is generated by summing the squares of the harmonic amplitudes within each zone. The total harmonic distortion and the zone harmonic energy are recorded in time series.
6. The method for monitoring and alerting power usage as claimed in claim 1, characterized in that, The method uses low-frequency metering channel sampling data, total harmonic distortion, and zonal harmonic energy as exogenous inputs to construct an improved structured state-space sequence model, outputting future multi-step prediction results and model hidden states, including: Based on low-frequency metering channel sampling data and total harmonic distortion and zonal harmonic energy as exogenous inputs, an improved structured state-space sequence model is constructed, which includes a harmonic state transition module, a harmonic gated decoding module, and a multi-task output module. In the harmonic state transition module, time-varying state transition parameters and exogenous input mapping parameters are generated using the exogenous input vector and the initial hidden state of the model as variables to update the hidden state of the model. The exogenous input vector is composed of the low-frequency metering channel sampling data and the total harmonic distortion and the zonal harmonic energy at the corresponding time step. The initial hidden state of the model is calculated by the initial state mapping unit using the feature vector of the low-frequency metering channel sampling data at the first analysis time step and the total harmonic distortion and the zonal harmonic energy at the corresponding time step. In the harmonic gated decoding module, the updated model hidden state is connected with the total harmonic distortion and the zonal harmonic energy feature vectors and then input into the gated function to generate gated coefficients. Based on the gated coefficients, the updated model hidden state is gated and modulated to obtain the gated and modulated hidden state representation. In the multi-task output module, the hidden state representation and the exogenous input vector are input together for processing to generate multi-step prediction results for the future, and the predicted values corresponding to each prediction time step are output respectively. Each prediction time step corresponds to an independent readout parameter, and the number of prediction steps is determined by the preset prediction range. The gating coefficients, time-varying state transition parameters, and exogenous input mapping parameters are recorded, and the prediction results for future multiple steps and the updated model hidden states are output and cached in chronological order.
7. The method for monitoring and alerting power usage as claimed in claim 1, characterized in that, The process of gating and fusing the rotating harmonic state sequence with the model's latent states, applying time-domain and frequency-domain consistency constraints to the future multi-step prediction results, and generating a search space that defines the harmonic parameter set by the prior ranges of harmonic frequency and amplitude includes: The generated gating coefficients are used to perform element-wise weighted fusion of the model's hidden states to form a gated fused hidden state representation. The current prediction sequence is extracted from the multi-step prediction results in the future. At the same time, the corresponding harmonic amplitude, harmonic frequency and harmonic phase parameters are extracted from the rotating harmonic state sequence at the current moment. The reference current sequence is synthesized in the time domain, and the sum of the squared differences between the current prediction sequence and the reference current sequence in multiple future time steps is calculated to obtain the time domain consistency constraint. Discrete Fourier transforms are performed on the current prediction sequence and the reference current sequence respectively to obtain the prediction spectrum and the reference spectrum. Within the frequency neighborhood range centered on each target harmonic frequency and determined by the preset frequency neighborhood width, the sum of the absolute values of the differences between the prediction spectrum and the reference spectrum is calculated to obtain the frequency domain consistency constraint. Based on the gated fusion hidden state representation, harmonic frequency prior values and harmonic amplitude prior values are generated, and a center value and positive and negative range are set for each harmonic frequency prior value and harmonic amplitude prior value to form frequency prior interval and amplitude prior interval, thus limiting the search space of the harmonic parameter set; The gating coefficients, gating fusion hidden state representations, time-domain consistency constraints, frequency-domain consistency constraints, frequency prior intervals, and amplitude prior intervals are recorded and saved together with the future multi-step prediction results.
8. A power usage status reminder and safety warning monitoring system, used to implement the power usage status reminder and safety warning monitoring method according to any one of claims 1 to 7, characterized in that, Includes the following modules: The acquisition and preprocessing module is used to acquire sampling data from the low-frequency metering channel and the high-frequency harmonic channel, and to preprocess the sampling data from the high-frequency harmonic channel to generate a high-frequency harmonic analysis window signal. The harmonic parameter module is used to perform parametric spectrum estimation based on the high-frequency harmonic analysis window signal and to filter effective harmonic components. The state sequence construction module is used to generate a rotating harmonic state sequence based on the harmonic parameter set, perform time smoothing on adjacent analysis windows, and calculate the total harmonic distortion and zonal harmonic energy. The state-space modeling module is used to build an improved structured state-space sequence model and output the prediction results for future multiple steps and the model's hidden states. The fusion constraint module is used to gating the fusion of the rotating harmonic state sequence and the model's hidden state, and to apply time-domain consistency and frequency-domain consistency constraints to the future multi-step prediction results. The joint optimization module is used to calculate the prediction error based on the future multi-step prediction results and the actual measurement sequence, construct the joint optimization objective, and solidify the structured state-space sequence model structure and fusion layer parameters. The monitoring and early warning module is used to deploy the structured state-space sequence model to the monitoring platform, trigger prompts and early warnings, and output them to the visualization terminal.
Citation Information
Patent Citations
Time series data trend prediction method, system and related device
CN113761022A
Multi-dimensional time sequence photovoltaic grid-connected harmonic current prediction control method
CN119496136A
Method and system for optimizing intelligent ice melting device of ultra-high voltage power transmission line
CN120197529A
Intelligent power security situation assessment method based on multi-source sensor network
CN120430635A
Waveform distortion mitigation in power systems
US20130038123A1
