A power grid voltage long-time series dynamic prediction method and system

Abstract

The application discloses a power grid voltage long-time sequence dynamic prediction method and system, and the method comprises the following steps: obtaining historical voltage effective value data and cleaning; using a signal decomposition algorithm to separate the data into a trend component and a fluctuation component; extracting global long-time sequence features and local mutation features through a double-flow network; using an adaptive gating unit to fuse the features and inputting the decoder to output voltage prediction values and fluctuation intervals; and generating a stability evaluation index based on the fluctuation interval. Through multi-scale feature decoupling and physical constraint loss function optimization, the application can accurately capture the long-time evolution and instantaneous disturbance of the voltage, solve the problems of lagging prediction of traditional models for mutation working conditions and lack of physical consistency, and provide a reliable basis for power grid safety scheduling.

Description

Technical Field

[0001] This invention belongs to the field of smart grid technology, and specifically relates to a method and system for long-term dynamic prediction of grid voltage. Background Technology

[0002] With the rapid development of new power systems and the integration of high proportions of distributed renewable energy and impulsive loads, the spatiotemporal distribution characteristics of grid voltage have become increasingly complex. Voltage is not only a key indicator for measuring power quality but also a core element for maintaining the safe and stable operation of the power grid. Therefore, accurate and real-time prediction of future voltage trends is of great significance for grid dispatching, reactive power optimization, and relay protection. Currently, grid voltage prediction techniques are mainly divided into physical modeling methods and data-driven methods. Due to the complexity of grid topology and the time-varying parameters, pure physical modeling is difficult to meet real-time requirements; therefore, data-driven methods based on artificial intelligence have gradually become a research hotspot.

[0003] Chinese invention patent application CN121146187A proposes a low-voltage prediction method for distribution network areas based on GA-BP and LSTM. This method first uses a GA-BP model to predict load, then uses the predicted load data as input to predict voltage through an LSTM model, dynamically adjusting the voltage threshold based on seasonal and meteorological conditions. However, this type of method has the following limitations: it is typically based on 15-minute or 1-hour data from an electricity consumption information collection system. While this long-interval data can reflect macro trends, it smooths out a large amount of critical transient fluctuations in the power grid. For voltage sags, flicker, or short-term oscillations on a second-scale basis, such models cannot effectively capture them, resulting in significant lag in the prediction results. Furthermore, this method directly inputs the raw data into the LSTM network without physically decoupling the long-term trend and instantaneous fluctuations in the voltage signal. During periods of severe voltage fluctuations, the LSTM is susceptible to high-frequency noise interference, leading to misjudgments of the overall trend.

[0004] Chinese invention patent application CN119940616A proposes a method for predicting node voltage in low-voltage distribution networks based on a full-element learning graph neural network (GNN). This method focuses on addressing the problem of missing topology features by iteratively updating edge and node features through a GNN and inferring voltage distribution using spatial correlation. However, this method still faces challenges in practical applications: GNN methods primarily mine the spatial topological relationships between nodes, but fail to adequately explore the complex evolution of single-node voltage signals in the time-frequency domain (such as the coexistence of long-term memory and short-term abrupt changes). Most such models only aim to minimize the mathematical mean square error during training, which may result in predicted values ​​that, while numerically small, violate the physical safety constraints of the power grid. Existing technologies mostly output a single voltage prediction value, failing to provide a confidence interval or fluctuation range for the prediction result. In complex power grid environments, dispatchers need to know the upper and lower bounds of potential voltage fluctuations to assess the potential risks to the system.

[0005] In summary, existing voltage prediction methods struggle to simultaneously assess both the accuracy of long-term trends and the sensitivity to short-term abrupt changes. Therefore, there is an urgent need for a long-term dynamic prediction method for grid voltage that can integrate multi-scale time-frequency characteristics, possess physical constraints, and provide fluctuation range assessment. Summary of the Invention

[0006] This invention provides a method and system for long-term dynamic prediction of grid voltage. By using adaptive signal decomposition based on sample entropy, extraction of time-frequency dual-stream features, and optimization of a composite loss function that introduces physical limit penalties, it aims to solve the problems of existing technologies in processing mesoscale voltage data, such as difficulty in taking into account both long-term trends and instantaneous fluctuations, lack of physical consistency constraints in prediction results, and inability to quantify and assess voltage stability risks in real time.

[0007] To address the aforementioned technical problems, this invention proposes a long-term dynamic prediction method for grid voltage, comprising the following steps:

[0008] Acquire historical voltage RMS time-series data from power grid monitoring nodes, and perform data cleaning and standardization.

[0009] Based on the signal decomposition algorithm, the processed time series data is separated into trend components that reflect long-term change patterns and fluctuation components that reflect instantaneous disturbances.

[0010] Feature extraction is performed on the trend component and the fluctuation component respectively to obtain global long-term time-series features characterizing the voltage evolution law, and local abrupt change features characterizing the instantaneous voltage fluctuation characteristics.

[0011] The global long-term features and local abrupt change features are weighted and fused using an adaptive gating unit to generate a multimodal fusion feature vector, which is then input into the prediction decoder to obtain the voltage prediction value and voltage fluctuation range at future time.

[0012] Based on the comparison between the predicted voltage fluctuation range and the power grid safety threshold, a voltage stability assessment index is generated.

[0013] Preferably, the signal decomposition algorithm employs a variational mode decomposition algorithm based on sample entropy optimization; the specific process includes:

[0014] Calculate the sample entropy of the original voltage RMS time series data, and determine the number of modes K based on the magnitude of the sample entropy;

[0015] When the sample entropy is greater than the preset threshold, increase the value of K; when the sample entropy is less than the preset threshold, decrease the value of K.

[0016] The modal component with the lowest center frequency is reconstructed as the trend component, and the remaining modal components are superimposed and reconstructed as the fluctuation component.

[0017] Preferably, the method for obtaining the global long-term features and local mutation features is as follows:

[0018] Construct a two-stream feature extraction network that includes a first extraction branch and a second extraction branch in parallel;

[0019] The trend component is input into the first extraction branch, and the attention mechanism is used to capture long-distance dependencies, and the global long-time features are output.

[0020] The fluctuation component is input into the second extraction branch, and a multi-scale convolutional neural network is used to capture local details at different time spans, and the local mutation features are output.

[0021] Preferably, the method for extracting the global long-term temporal features is as follows:

[0022] The trend component is processed based on a probabilistic sparse self-attention mechanism;

[0023] Specifically, this includes: calculating the attention score distribution between query vectors and key vectors, and selecting the top-u dominant query vectors that contribute the most to the trend of preceding voltage changes based on KL divergence;

[0024] The dot product operation and context aggregation are performed on the dominant query vector to generate global long-term features.

[0025] Preferably, the method for extracting the local mutation features is as follows:

[0026] The fluctuation components are processed using a multi-scale causal dilated convolutional network.

[0027] The convolutional network contains at least three parallel convolutional channels, and the kernel dilation rate d of each channel is determined according to... Exponential growth, where i is the layer index;

[0028] Setting causal padding ensures that when extracting voltage fluctuation characteristics at the current time t, only historical information from time t and earlier is used;

[0029] The feature maps output from each channel are stitched together to generate local mutation features containing receptive fields of different time spans.

[0030] Preferably, the adaptive gating unit uses the following formula for weighted fusion:

[0031]

[0032] In the formula, For the weighted fusion result, It is a global long-term feature. This is a characteristic of local mutations;

[0033] Fusion weights The generation method is as follows: global long-term features With local mutation characteristics After concatenation, it is mapped through a fully connected layer and calculated using the Sigmoid activation function. .

[0034] Preferably, the training process of the predictive decoder employs a physically constrained composite loss function, which includes a voltage limit violation penalty term;

[0035] The voltage over-limit penalty term is constructed as follows:

[0036] Set upper and lower threshold values ​​for safe operation of the power grid voltage; during training, when the voltage value predicted by the model falls within the upper and lower threshold values, the penalty term is zero;

[0037] When the voltage value predicted by the model exceeds the upper and lower thresholds, the magnitude of the excess is calculated using the ReLU activation function, and an exponentially increasing penalty weight is applied.

[0038] Preferably, the voltage stability evaluation index is generated in the following way:

[0039] The width of the voltage fluctuation range within the future prediction time window is calculated, and the voltage stability margin index is calculated in combination with the energy value of the fluctuation component; when the voltage stability margin index is lower than a preset threshold, the current power grid state is determined to be a critical instability state.

[0040] Preferably, the sampling time interval for the historical voltage effective value time series data is 1 to 5 seconds.

[0041] A second aspect of the present invention also provides a long-term dynamic prediction system for grid voltage, the system being used to implement the method as described in the first aspect of the present invention, comprising:

[0042] The data acquisition and preprocessing module is used to acquire historical voltage RMS time-series data of power grid monitoring nodes and perform data cleaning and standardization.

[0043] The signal decomposition module uses a signal decomposition algorithm to separate the processed voltage RMS time series data into a trend component that reflects long-term variation patterns and a fluctuation component that reflects instantaneous disturbances.

[0044] The dual-stream feature extraction module is used to extract features from the trend component and the fluctuation component respectively, to obtain global long-time-series features characterizing the voltage evolution law, and local abrupt change features characterizing the instantaneous voltage fluctuation characteristics.

[0045] The adaptive fusion module uses an adaptive gating unit to perform weighted fusion of the global long-term features and local mutation features to generate a multimodal fusion feature vector;

[0046] The prediction decoding module is used to input the fused feature vector into the prediction decoder and output the voltage prediction value and voltage fluctuation range at future time.

[0047] The stability assessment module generates voltage stability assessment indicators based on the comparison between the predicted voltage fluctuation range and the power grid safety threshold.

[0048] Compared with the prior art, the present invention has the following technical effects:

[0049] 1. The prediction method proposed in this invention utilizes variational mode decomposition technology based on sample entropy optimization to physically decouple complex voltage signals into long-term trend components and instantaneous fluctuation components. Combined with a dual-stream network structure of attention mechanism + multi-scale dilated convolution, it can simultaneously and accurately capture the daily cycle evolution of voltage as well as the instantaneous perturbation details at the microsecond to second level, significantly improving the model's prediction sensitivity for sudden changes in voltage conditions such as voltage dips and surges.

[0050] 2. The prediction method proposed in this invention introduces a composite loss function that includes a voltage limit penalty term, which internalizes the physical constraints of power grid safe operation into the training process of the deep learning model, effectively avoiding the generation of prediction values ​​that violate physical laws by the pure data-driven model; at the same time, the voltage fluctuation range output by the model provides dispatchers with a quantitative basis for uncertainty risk assessment, making up for the deficiency of traditional point prediction in reflecting the fluctuation range.

[0051] 3. The prediction method proposed in this invention, through the design of an adaptive gating fusion unit, enables the model to automatically sense the grid operating status based on the current voltage characteristics and dynamically adjust the weight allocation of trend features and fluctuation features, thereby maintaining extremely high prediction robustness under different load characteristics and complex grid environments.

[0052] 4. The prediction method proposed in this invention constructs a voltage stability index that integrates the prediction interval width and fluctuation energy, which can transform complex time-series prediction data into an intuitive power grid health status level, providing decision support with rapid response capabilities for the power grid's proactive defense, load shedding, and reactive power dispatch. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the prediction method described in this invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0055] Example 1

[0056] This embodiment presents a method for long-term dynamic prediction of grid voltage, such as... Figure 1 As shown, it includes the following steps one through five:

[0057] Step 1: Obtain historical voltage RMS time-series data of power grid monitoring nodes, and perform data cleaning and standardization.

[0058] First, voltage signals are collected by voltage monitoring devices deployed at key nodes of the power grid. These key nodes include substation busbars, the low-voltage side of distribution transformers, or critical load terminals. Voltage monitoring devices include synchronous phasor measurement units (PMUs), smart meters, or SCADA (Supervisory Control and Data Acquisition) system terminals.

[0059] The sampling time interval for the historical voltage RMS time-series data is 1–5 seconds. In this embodiment, the sampling time interval is set to... .

[0060] Setting a sampling interval of 1 to 5 seconds significantly reduces data storage and computational burden compared to microsecond-level waveform data. Compared to traditional 15-minute-level SCADA data, this granularity can preserve short-term voltage fluctuation characteristics, such as sags caused by motor startup and fluctuations caused by photovoltaic cloud shading, thus balancing the real-time nature of prediction with the ability to capture long-term trends.

[0061] The acquired raw voltage sequence is denoted as ,in Let N represent the effective voltage value at the t-th sampling time, and N be the total length of the sequence.

[0062] Due to packet loss or sensor malfunction, the original sequence V may contain missing or outlier values, requiring cleaning.

[0063] Missing value handling: detect null values ​​or zero values ​​in the sequence under non-power outage conditions. If the number of consecutive missing points is less than the preset threshold (e.g., 3 points), linear interpolation is used to fill in the missing points. If the number of consecutive missing points is too large, the sample segment is removed.

[0064] The linear interpolation formula is:

[0065]

[0066] in, The missing voltage value to be filled; , These are the most recent effective voltage observations before and after the missing point, respectively. , , These are the corresponding timestamps.

[0067] Outlier handling employs the Laida criterion (3). (Criteria) Identify noise caused by measurement errors. Calculate the mean within the sliding window. and standard deviation If the voltage value at a certain moment t satisfy If the value is not found to be an outlier, it will be corrected using the linear interpolation method described above to generate a new value. To prevent noise from interfering with the model's learning of the true fluctuation trend; for non-outliers If the value is 0, then its original value is retained.

[0068] To eliminate the difference in data units between different voltage levels and accelerate the convergence speed of the neural network model, the cleaned voltage data is subjected to Z-Score standardization. The standardization formula is as follows:

[0069]

[0070] in, The standardized voltage at time t is input as a characteristic value; The original effective voltage value at time t after cleaning; The average voltage value in the historical training dataset; is the voltage standard deviation of the historical training dataset.

[0071] To meet the input requirements of supervised learning models, a sliding window technique is used to divide the standardized long sequence into multiple sample pairs. :

[0072] The input sequence X contains the past Historical data at each time step, i.e. ;

[0073] Tag sequence Includes the future The actual voltage value at each time step, i.e. .

[0074] In this embodiment, the input window length is set. (i.e., data from the past 6 minutes, based on 3-second sampling), prediction step size (i.e., data from the next minute) to construct a standard dataset for model training in subsequent steps.

[0075] Step 2: Based on the signal decomposition algorithm, the processed time series data is separated into trend components that reflect long-term change patterns and fluctuation components that reflect instantaneous disturbances.

[0076] In this embodiment, to address the non-stationary and nonlinear characteristics of the power grid voltage signal, as well as the differences in voltage fluctuation complexity under different load conditions, the signal decomposition algorithm employs a variational mode decomposition (VMD) algorithm based on sample entropy optimization; the specific process includes:

[0077] First, the voltage input sequence after standardization in step one... Complexity assessment is performed. The sample entropy of the original voltage RMS time series data is calculated, and the number of modes K is determined based on the magnitude of the sample entropy. The sample entropy can measure the probability of new modes emerging in the voltage time series. The larger the entropy value, the higher the randomness and complexity of the voltage fluctuations; the smaller the entropy value, the more stable the voltage changes or the obvious periodicity.

[0078] Sample Entropy The calculation formula is as follows:

[0079]

[0080] in, That is, sample entropy m is the embedding dimension, representing the length of the vector used to reconstruct the phase space; in this embodiment, m=2. r is the similarity tolerance; in this embodiment, r is 0.2 times the standard deviation of the data. N is the data length of the voltage sequence. This represents the probability that the distance between two sets of voltage vectors is less than the tolerance r in m-dimensional space; This represents the probability that the distance between two sets of voltage vectors is less than the tolerance r in the m+1 dimensional space.

[0081] To avoid the under-decomposition or over-decomposition problems caused by a fixed K value in traditional VMD algorithms, this embodiment establishes a mapping relationship between sample entropy and the K value: when the sample entropy is greater than a preset threshold, the K value is increased; when the sample entropy is less than the preset threshold, the K value is decreased. As an example of K value mapping, a specific implementation strategy in this embodiment is as follows:

[0082] Set the reference mode number and complexity threshold range In one example of this embodiment, the complexity threshold range is [0.2, 0.6]:

[0083] like This indicates that the voltage is in a stable period, and the setting... ;

[0084] like This indicates that the voltage is within the normal fluctuation period. ;

[0085] like This indicates that the voltage is in a period of severe fluctuation. (Setting) (or higher, depending on the specific power grid environment).

[0086] Using a defined number of modes K, variational mode decomposition is performed on the input voltage sequence. The original signal is decomposed into K modes with a center frequency of K. The band-limited eigenmode function centered at the center is denoted as By introducing a quadratic penalty term and Lagrange multipliers, the variational problem is solved iteratively using the alternating direction multiplier method, ultimately yielding K modal components.

[0087] After obtaining K modal components Then, based on its center frequency Arrange them in ascending order. Reconstruct the trend component from the modal component with the lowest center frequency, and superimpose the remaining modal components to reconstruct the fluctuation component.

[0088] The specific calculation formula is as follows:

[0089] Trend Components The mode with the smallest center frequency was selected as the trend component, reflecting the slow voltage drift caused by the daily load variation curve of the power grid. .

[0090] Fluctuation components The remaining high-frequency modes are superimposed as fluctuation components to reflect instantaneous voltage disturbances caused by equipment start-up and shutdown, short-circuit faults, or the randomness of new energy output. .

[0091] Step 3: Extract features from the trend component and the fluctuation component respectively to obtain global long-term time-series features characterizing the voltage evolution law and local abrupt change features characterizing the instantaneous voltage fluctuation characteristics.

[0092] To balance the long-term dependence and short-term abrupt changes in voltage variations, the specific method for obtaining the global long-term time-series features and local abrupt change features is as follows:

[0093] Construct a two-stream feature extraction network that includes a first extraction branch and a second extraction branch in parallel;

[0094] The trend component is input into the first extraction branch, and the attention mechanism is used to capture long-distance dependencies, and the global long-time features are output.

[0095] The fluctuation component is input into the second extraction branch, and a multi-scale convolutional neural network is used to capture local details at different time spans, and the local mutation features are output.

[0096] The specific implementation process includes the first extraction branch performing global long-term feature extraction, and the second extraction branch performing local mutation feature extraction.

[0097] For the smoothed trend component separated in step two The improved Transformer architecture is used for processing, and the method for extracting the global long-term temporal features is as follows:

[0098] The trend component is processed based on a probabilistic sparse self-attention mechanism;

[0099] Specifically, this includes: calculating the attention score distribution between query vectors and key vectors, and selecting the top-u dominant query vectors that contribute the most to the preceding voltage change trend based on KL divergence (Kullback-Leibler Divergence);

[0100] The dot product operation and context aggregation are performed on the dominant query vector to generate global long-term features.

[0101] Specifically, the input trend component sequence is first mapped to a high-dimensional feature vector. Let the length of the input sequence be L, and the feature dimension be... The query matrix Q, key matrix K, and value matrix V are generated through linear projection, where... .

[0102] To reduce the computational complexity of traditional attention mechanisms, KL divergence is used to measure the i-th query vector. The difference between the attention distribution and the uniform distribution. The greater the difference, the more important the information that the query vector focuses on. Sparsity measurement formula. Defined as:

[0103]

[0104] in, To query the vector in the i-th row of matrix Q; Let be the j-th row vector in the key matrix K; d is the scaling factor. In this embodiment... .

[0105] according to Sort all query vectors by their size, select the top u dominant query vectors, and form a sparse query matrix. .in , where c is a hyperparameter. Only calculate... The output features are obtained by dot product attention with K. :

[0106]

[0107] The final result This is a global long-term feature, which effectively captures the evolution of voltage over long periods and has significantly better computational efficiency than the full attention mechanism.

[0108] Regarding the fluctuation components separated in step two It contains high-frequency noise and instantaneous drop features, requiring precise capture of local patterns. The method for extracting the local abrupt change features is as follows:

[0109] The fluctuation components are processed using a multi-scale causal dilated convolutional network.

[0110] The convolutional network contains at least three parallel convolutional channels, and the kernel dilation rate d of each channel is determined according to... Exponential growth, where i is the layer index;

[0111] Setting causal padding ensures that when extracting voltage fluctuation characteristics at the current time t, only historical information from time t and earlier is used;

[0112] The feature maps output from each channel are stitched together to generate local mutation features containing receptive fields of different time spans.

[0113] Specifically, first, a convolutional module containing M parallel channels (or layers) is constructed, and the index is... For the i-th channel, the kernel size is set to... ,For example The expansion rate is The formula for dilated convolution is as follows:

[0114]

[0115] in, The output characteristics of the i-th channel at time t; For the input fluctuation component data; Let j be the weight parameter of the i-th channel convolution kernel; This is a bias term.

[0116] This embodiment performs zero-padding on the left side (past time steps) of the input sequence before convolution. Padding length Depends on the expansion rate and convolution kernel Size:

[0117]

[0118] This ensures that when calculating the convolution output at time t, the receptive field of the convolution kernel only covers time t and the time before t.

[0119] Different expansion rates Corresponding to different receptive fields; when i=0 and d=1, the receptive field is small, capturing high-frequency minute jitters; when i=2 and d=4, the receptive field is large, capturing voltage dips or oscillations lasting several seconds.

[0120] Finally, output all channels Perform channel-level stitching:

[0121]

[0122] Received This is a local mutation feature that preserves the transient details of the voltage signal at different time scales.

[0123] Step 4: Use an adaptive gating unit to perform weighted fusion of the global long-term features and local abrupt change features to generate a multimodal fusion feature vector, which is then input into the prediction decoder to obtain the voltage prediction value and voltage fluctuation range at future time points.

[0124] In this embodiment, considering that the dominant characteristics of the grid voltage differ under different operating conditions (such as the stable operation period and the fault disturbance period), in order to achieve dynamic optimization of the characteristics, the adaptive gating unit uses the following formula for weighted fusion:

[0125]

[0126] In the formula, For the weighted fusion result, It is a global long-term feature. This is a characteristic of local mutations;

[0127] Fusion weights It is a scalar or vector that changes dynamically over time, used to measure whether the model should trust long-term trends or instantaneous fluctuations at the current moment. Fusion weights The generation method is as follows: global long-term features With local mutation characteristics After concatenation, it is mapped through a fully connected layer and calculated using the Sigmoid activation function. .

[0128] The specific calculation process is as follows:

[0129] First, in terms of feature dimensions and By concatenating the features, a joint feature vector is obtained. :

[0130]

[0131] Then, weights are generated through a learnable linear mapping layer and a sigmoid non-linear activation function:

[0132]

[0133] In the above formula, This is the weight matrix of the gated unit; The bias vector of the gated unit; This indicates a vector concatenation operation.

[0134] When the power grid is in a phase of stable load changes, the model will automatically learn to... Approaching 1, with emphasis on When a power grid fault occurs or a large motor starts, causing a sudden voltage drop, It will approach 0, focusing on This enables precise capture of voltage dynamic characteristics.

[0135] The fused features The input is a predictive decoder. In this embodiment, the decoder employs a Multi-Layer Perceptron (MLP) or Long Short-Term Memory (LSTM) network. Its output comprises two parts: point prediction values ​​and interval prediction values. The point prediction value corresponds to the expected value of the voltage at a future time. The interval prediction value corresponds to the upper and lower bounds of the voltage at future times (e.g., quantiles). and This allows us to determine the voltage fluctuation range.

[0136] To address the issue that purely data-driven models are prone to predicting violations of power grid physical laws (such as significant voltage exceedances without penalty), the training process of the prediction decoder employs a physically constrained composite loss function, which includes a voltage exceedance penalty term. Defined as:

[0137]

[0138] in, For mean square error loss, For quantile loss, This is a penalty item for voltage exceeding the limit. For hyperparameter weights.

[0139] The voltage over-limit penalty term is constructed as follows:

[0140] Set upper and lower threshold values ​​for safe operation of the power grid voltage; during training, when the voltage value predicted by the model falls within the upper and lower threshold values, the penalty term is zero;

[0141] When the voltage value predicted by the model exceeds the upper and lower thresholds, the magnitude of the excess is calculated using the ReLU activation function, and an exponentially increasing penalty weight is applied.

[0142] The specific calculation formula is as follows: Let the upper limit of safe operation of the power grid be... The lower limit is The model predicts the value. .

[0143] First, calculate the over-limit amplitude value. :

[0144]

[0145] The ReLU function is defined as follows: Ensure that only when the voltage is expected or hour, It is positive only if it is positive otherwise it is 0.

[0146] Next, construct an exponential penalty term. :

[0147]

[0148] in, Basic penalty coefficient; As an exponential growth factor, there is .

[0149] Once the predicted voltage exceeds the limit, the loss value will increase with the magnitude of the voltage exceeding the limit. The increase in [the number of parameters] is exponential and explosive. This forces the model to avoid producing out-of-bounds predictions when updating parameters through backpropagation, unless the input data actually contains strong fault characteristics, thus significantly improving the model's prediction reliability under critical conditions.

[0150] Step 5: Based on the comparison between the predicted voltage fluctuation range and the power grid safety threshold, generate voltage stability assessment indicators.

[0151] The voltage stability evaluation index is generated as follows:

[0152] First, based on the future output of step four The voltage prediction results at each time point (prediction time window) are calculated. The width of the voltage fluctuation range within the future prediction time window is calculated, and the voltage stability margin index (VSMI) is calculated by combining the energy value of the fluctuation component.

[0153] The upper limit of voltage at time t is predicted based on the quantile regression in step four. and lower limit Calculate the absolute width of the interval at time t. This is used to reflect the degree of uncertainty of voltage at future moments:

[0154]

[0155] Throughout the forecast time window Average fluctuation width within for:

[0156]

[0157] Using the wave components separated in step two Calculate its signal energy within the most recent historical window. This reflects the current level of "noise" in the power grid (such as the presence of persistent forced oscillations or high-frequency harmonics).

[0158]

[0159] in, The length of the historical input sequence.

[0160] VSMI is a normalized index used to comprehensively measure voltage offset risk and oscillation risk. This embodiment uses a negative exponential function to construct VSMI, ensuring its value falls within a certain range. between:

[0161]

[0162] in, The rated voltage of the grid node is used for normalization; This is the weighting coefficient for the fluctuation width, representing the sensitivity to forecast uncertainty; an example value is 0.6. This is the weighting coefficient for the energy value, representing the sensitivity to the current oscillation intensity; an example value is 0.4. This represents an exponential function with the natural constant e as its base.

[0163] When the power grid is operating smoothly, the prediction range is narrow ( Low fluctuation energy ( At this point, the VSMI approaches 1, indicating system stability. When a grid fault or severe fluctuation occurs, the prediction interval widens and the fluctuation energy surges, causing the exponential term to increase, resulting in a rapid decrease in the VSMI approaching 0.

[0164] After obtaining the VSMI, it is compared with a preset safety threshold in real time to determine the current operating state level of the power grid. When the voltage stability margin index is lower than the preset threshold, the current power grid state is determined to be a critical instability state.

[0165] The specific tiered assessment strategy is as follows:

[0166] Set two key thresholds: warning threshold (e.g., 0.7) and critical threshold (For example, 0.4). If The system determines that the grid voltage is in a healthy state and requires no intervention. If... If a minor disturbance or trending voltage shift is detected in the power grid, and the system generates a yellow alert, the dispatcher is advised to monitor the switching status of reactive power compensation devices. If the power grid is determined to be in a critical unstable state, it indicates that the voltage is highly likely to exceed the limit or oscillate continuously in the near future. If the system triggers a red alarm, it will transmit the alarm signal and related predictive characteristic parameters to the power grid dispatch and control center and initiate the intervention process.

[0167] In a preferred embodiment of the present invention, a sixth step (intervention process) is further included: based on the voltage stability assessment index and the predicted voltage fluctuation range, a graded power grid voltage intervention strategy is formulated and implemented.

[0168] Specifically, when the power grid is determined to be in a stable state, the system does not perform forced intervention, but sends a fine-tuning command to the Automatic Voltage Control (AVC) system based on the future voltage fluctuation range output in step four. If the predicted voltage fluctuation range shows a slow upward trend and is close to the upper limit, the system instructs the tap changer of the on-load tap-changing transformer to be lowered by one tap in advance to reserve a safety margin.

[0169] When the system issues a yellow alert, the rapid response mechanism of the reactive power compensation device is activated. Based on the predicted voltage drop, the required reactive power support is calculated, and the static var generator or parallel capacitor bank is instructed to switch on or off to prioritize the smoothing of local voltage fluctuations and prevent further deterioration of the situation.

[0170] When the system triggers a red alert, it indicates a high risk of voltage collapse, requiring the implementation of emergency control strategies such as:

[0171] If a line fault is detected and reclosing may exacerbate voltage oscillations, the reclosing function of the relevant circuit breaker will be temporarily disabled.

[0172] Based on the energy value of the fluctuation component extracted in step two, the feeder area where the disturbance source is located is identified, and 10-20% of the non-critical load in this area is precisely cut off to block the chain reaction of voltage instability.

[0173] Send a strong excitation command to the excitation system of the critical generator to increase the terminal voltage and maintain the voltage stability level of the system.

[0174] Example 2

[0175] This embodiment is a long-term dynamic prediction system for grid voltage. The system is used to implement the method described in Embodiment 1, including:

[0176] The data acquisition and preprocessing module is used to acquire historical voltage RMS time-series data of power grid monitoring nodes and perform data cleaning and standardization.

[0177] The signal decomposition module uses a signal decomposition algorithm to separate the processed voltage RMS time series data into a trend component that reflects long-term variation patterns and a fluctuation component that reflects instantaneous disturbances.

[0178] The dual-stream feature extraction module is used to extract features from the trend component and the fluctuation component respectively, to obtain global long-time-series features characterizing the voltage evolution law, and local abrupt change features characterizing the instantaneous voltage fluctuation characteristics.

[0179] The adaptive fusion module uses an adaptive gating unit to perform weighted fusion of the global long-term features and local mutation features to generate a multimodal fusion feature vector;

[0180] The prediction decoding module is used to input the fused feature vector into the prediction decoder and output the voltage prediction value and voltage fluctuation range at future time.

[0181] The stability assessment module generates voltage stability assessment indicators based on the comparison between the predicted voltage fluctuation range and the power grid safety threshold.

[0182] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for long-term dynamic prediction of grid voltage, characterized in that, Includes the following steps: Acquire historical voltage RMS time-series data from power grid monitoring nodes, and perform data cleaning and standardization. The processed time-series data is separated into a trend component reflecting long-term changes and a fluctuation component reflecting instantaneous disturbances based on a signal decomposition algorithm. The signal decomposition algorithm adopts a variational mode decomposition algorithm based on sample entropy optimization. The specific process is as follows: calculate the sample entropy of the original voltage RMS time-series data, and determine the number of modes K according to the magnitude of the sample entropy; when the sample entropy is greater than a preset threshold, increase the value of K; when the sample entropy is less than the preset threshold, decrease the value of K; reconstruct the trend component with the lowest center frequency, and superimpose the remaining mode components to reconstruct the fluctuation component. A dual-stream feature extraction network containing a first extraction branch and a second extraction branch in parallel is constructed. The trend component is input into the first extraction branch, and the long-distance dependency is captured by the attention mechanism. The trend component is processed based on the probabilistic sparse self-attention mechanism. The attention score distribution of the query vector and the key vector is calculated. The top-u dominant query vectors with the highest contribution to the preceding voltage change trend are selected based on KL divergence. The dominant query vectors are subjected to dot product operation and context aggregation to output global long-time features that characterize the voltage evolution law. The fluctuation component is input into the second extraction branch, and a multi-scale convolutional neural network is used to capture local details across different time spans. Specifically, a multi-scale causal dilated convolutional network is used to process the fluctuation component. The convolutional network contains at least three parallel convolutional channels, and the kernel dilation rate d of each channel is determined according to... Exponential growth, where i is the layer index, causal padding is set so that when extracting the voltage fluctuation features at the current time t, only the historical information at time t and before is used, the feature maps output by each channel are stitched together to generate local mutation features that contain receptive fields of different time spans and characterize the instantaneous voltage fluctuation characteristics. The global long-term features and local abrupt change features are weighted and fused using an adaptive gating unit to generate a multimodal fusion feature vector, which is then input into the prediction decoder to obtain the voltage prediction value and voltage fluctuation range at future time. Based on the comparison between the predicted voltage fluctuation range and the power grid safety threshold, a voltage stability assessment index is generated.

2. The method according to claim 1, characterized in that, The adaptive gating unit uses the following formula for weighted fusion: In the formula, For the weighted fusion result, It is a global long-term feature. This is a characteristic of local mutations; Fusion weights The generation method is as follows: global long-term features With local mutation characteristics After concatenation, it is mapped through a fully connected layer and calculated using the Sigmoid activation function. .

3. The method according to claim 1, characterized in that, The training process of the prediction decoder uses a physical constraint composite loss function, which includes a voltage limit violation penalty term. The voltage over-limit penalty term is constructed as follows: Set upper and lower threshold values ​​for safe operation of the power grid voltage; during training, when the voltage value predicted by the model falls within the upper and lower threshold values, the penalty term is zero; When the voltage value predicted by the model exceeds the upper and lower thresholds, the magnitude of the excess is calculated using the ReLU activation function, and an exponentially increasing penalty weight is applied.

4. The method according to claim 1, characterized in that, The voltage stability evaluation index is generated as follows: The width of the voltage fluctuation range within the future prediction time window is calculated, and the voltage stability margin index is calculated in combination with the energy value of the fluctuation component; when the voltage stability margin index is lower than a preset threshold, the current power grid state is determined to be a critical instability state.

5. The method according to claim 1, characterized in that, The sampling time interval for the historical voltage RMS time series data is 1 to 5 seconds.

6. A long-term dynamic prediction system for grid voltage, characterized in that, The system is used to implement the method as described in any one of claims 1-5, comprising: The data acquisition and preprocessing module is used to acquire historical voltage RMS time-series data of power grid monitoring nodes and perform data cleaning and standardization. The signal decomposition module uses a signal decomposition algorithm to separate the processed voltage RMS time series data into a trend component that reflects long-term variation patterns and a fluctuation component that reflects instantaneous disturbances. The dual-stream feature extraction module is used to extract features from the trend component and the fluctuation component respectively, to obtain global long-time-series features characterizing the voltage evolution law, and local abrupt change features characterizing the instantaneous voltage fluctuation characteristics. The adaptive fusion module uses an adaptive gating unit to perform weighted fusion of the global long-term features and local mutation features to generate a multimodal fusion feature vector; The prediction decoding module is used to input the fused feature vector into the prediction decoder and output the voltage prediction value and voltage fluctuation range at future time. The stability assessment module generates voltage stability assessment indicators based on the comparison between the predicted voltage fluctuation range and the power grid safety threshold.

Images