A phase modifier automatic start-stop control method based on energy consumption self-adaptive optimization, electronic equipment and storage medium

Abstract

The application discloses a phase modifier automatic start-stop control method based on energy consumption self-adaptive optimization, electronic equipment and a storage medium. The method comprises the following steps: acquiring multi-source operation data of the phase modifier and performing pretreatment and feature extraction, and constructing a standardized feature matrix; based on the matrix, a device dynamic energy consumption benchmark model is constructed by combining a long short-term memory network with a self-attention mechanism to predict a theoretical energy consumption baseline, and the theoretical energy consumption baseline is fused with power grid dispatching and load prediction data, and a future reactive power demand prediction sequence and a marginal energy consumption coefficient sequence are synchronously output through a time series convolution network; based on the above sequence, a multi-class energy consumption model containing start-stop transient, no-load and load operation loss is constructed, a dynamic start-stop decision boundary is generated through multi-objective optimization, and finally a start-stop control instruction resistant to disturbance is output by combining a sliding window comparison and fuzzy logic processing. The application realizes adaptive optimization of the phase modifier start-stop control, and improves the economy and reliability while ensuring the stability of the power grid.

Description

Technical Field

[0001] This invention relates to a synchronous condenser control method, and more particularly to an automatic start-stop control method, electronic device and storage medium for a synchronous condenser based on energy consumption adaptive optimization. Background Technology

[0002] With the continuous expansion of power system scale and the increasing proportion of renewable energy grid connection, the requirements for reactive power support capacity and voltage stability of the power grid are becoming increasingly stringent. Synchronous condensers, as important dynamic reactive power compensation devices, play a crucial role in maintaining grid voltage stability, suppressing voltage fluctuations, and improving system short-circuit capacity. Currently, the operation control of synchronous condensers is gradually evolving from traditional manual dispatching towards automation and intelligence, especially in start-up and shutdown control, where control methods based on rule logic, load forecasting, and dispatching commands are being introduced to achieve dynamic adjustment of equipment operating status. Simultaneously, with advancements in data acquisition and condition monitoring technologies, the large amount of multi-source data generated during the operation of synchronous condensers makes refined control possible, making data-driven optimization control methods a research hotspot.

[0003] However, existing automatic start-stop control methods for synchronous condensers still have the following significant drawbacks: 1. Existing methods mainly rely on fixed threshold strategies or single load forecast results for decision-making, lacking the ability to comprehensively model the energy consumption characteristics of equipment operation and multi-source operating conditions. This leads to start-stop decisions often being based solely on grid demand, ignoring the impact of equipment operating costs and state evolution on the decision. Due to the failure to effectively integrate historical operating data, equipment condition monitoring data, and environmental factors, existing technologies lack the ability to characterize the dynamic characteristics of synchronous condenser energy consumption changing with operating conditions, making it difficult to construct accurate energy consumption benchmark models, thus failing to achieve energy consumption-based adaptive optimization control. 2. When predicting future reactive power demand, traditional methods often use single time series models or empirical rules, lacking in-depth analysis of time-series dependencies and the coupling relationships between multi-source heterogeneous data, resulting in limited prediction accuracy and difficulty in supporting highly reliable start-stop decisions. 3. In the start-stop decision-making process, existing technologies typically do not comprehensively consider the trade-offs between start-stop transient losses, no-load operating losses, and load operating efficiency, lacking an effective multi-objective optimization mechanism. This makes it difficult to strike a balance between economic efficiency and equipment lifespan in decision-making, resulting in poor overall economic efficiency. 4. Existing methods lack effective decision smoothing and disturbance rejection mechanisms to address fluctuations and disturbances during real-time operation. They are prone to frequent start-ups, shutdowns, or malfunctions due to short-term demand fluctuations, affecting control stability and equipment safety. Summary of the Invention

[0004] Purpose of the invention: The first purpose of the present invention is to provide an automatic start-stop control method for a synchronous condenser based on adaptive energy consumption optimization, stable and reliable control commands, multi-objective optimization, and intelligent start-stop decision-making. The second purpose of the present invention is to provide an electronic device and a computer-readable storage medium for implementing the method.

[0005] Technical solution: The automatic start / stop control method for synchronous condensers based on adaptive energy consumption optimization described in this invention includes the following steps:

[0006] (1) Obtain historical operating data of synchronous condensers, power grid reactive power demand sequence, equipment status monitoring data and environmental parameters and perform preprocessing, extract statistical features, frequency domain features and cross features that reflect the internal state coupling of equipment, and construct a standardized multidimensional feature matrix;

[0007] (2) Based on the standardized multidimensional feature matrix, a dynamic energy consumption benchmark model for equipment is constructed to output the future theoretical energy consumption baseline; the future theoretical energy consumption baseline is fused with power grid dispatch plan data and load forecast data to generate a multi-source heterogeneous data fusion tensor; the multi-source heterogeneous data fusion tensor is input into a temporal convolutional network to obtain a sequence of reactive power demand forecasts and a sequence of marginal energy consumption coefficients for future periods; the dynamic energy consumption benchmark model for equipment is implemented using a long short-term memory network combined with a self-attention mechanism.

[0008] (3) Based on the reactive power demand prediction sequence and the marginal energy consumption coefficient sequence, construct a multi-type energy consumption assessment model including start-stop transient loss, no-load excitation loss and load operation loss, and generate start-stop decision boundary vectors corresponding to future time points through multi-objective optimization solution; compare the real-time monitored reactive power deficit with the start-stop decision boundary vectors within a sliding time window, and after fuzzy logic processing, finally output the synchronous condenser start-stop control command.

[0009] Preferably, in step (1), the preprocessing includes missing value imputation, outlier removal, and unified timescale alignment; the missing value imputation is performed using cubic spline interpolation; the outlier removal uses the Grubbs test to determine and remove outlier data points exceeding three times the standard deviation. This invention uses cubic spline interpolation to imput missing data. The main reason is that the sequence data such as voltage, current, temperature, and vibration generated during the operation of the synchronous condenser essentially reflect the continuous evolution of physical states. Especially when the equipment undergoes start-up and shutdown transitions or is in dynamic operating conditions such as deep phase advance or hysteresis, adjacent sampling points exhibit a smooth nonlinear change trend. Compared to linear interpolation, cubic spline interpolation can effectively preserve the local curvature information of the data sequence while avoiding the introduction of significant oscillations. This processing method provides a higher fidelity time-series data foundation for subsequent frequency domain feature analysis, thereby ensuring the accuracy of subsequent frequency domain feature extraction and avoiding feature distortion.

[0010] Preferably, in step (1), the cross-features include aging sensitivity features composed of insulation resistance and winding temperature, mechanical load coupling features composed of vibration amplitude and current fluctuation, temperature rise gradient features composed of stator winding hot spot temperature and ambient temperature, and apparent power features composed of the effective value of terminal line voltage and the effective value of stator line current.

[0011] Preferably, in step (2), the construction of the device dynamic energy consumption benchmark model includes the following steps:

[0012] Based on the standardized multidimensional feature matrix, the output vector of the long short-term memory network at each time step is obtained;

[0013] Calculate the self-attention weights and sum the output vectors at each time step to obtain the context vector;

[0014] The context vector is concatenated with the final time-end output of the long short-term memory network to form a comprehensive feature vector;

[0015] The integrated feature vector is mapped to the future theoretical energy consumption baseline.

[0016] Preferably, during the training phase of the device dynamic energy consumption benchmark model, the marginal energy consumption label for training the marginal energy consumption coefficient is constructed in the following manner:

[0017] Apply a preset fixed reactive power perturbation to the planned reactive power output values ​​in the sample;

[0018] Input the sample containing the planned reactive power output value after the perturbation into the trained dynamic energy consumption benchmark model of the equipment, and calculate the theoretical energy consumption baseline prediction value after the perturbation.

[0019] The marginal energy consumption label is constructed using the following formula:

[0020]

[0021] in, Labeling as marginal energy consumption; This is a preset fixed reactive power perturbation; This serves as the theoretical energy consumption baseline after the disturbance. Indicates the future of the original sample The actual energy consumption label value corresponding to each minute.

[0022] In the model training phase, this invention employs the finite difference method to construct a supervised label for the marginal energy consumption coefficient, transforming the "energy consumption change rate corresponding to a unit change in reactive power" into a numerical label that can be learned under supervision. This method avoids the need for actual field disturbance tests at every future moment, thus significantly improving the feasibility and engineering applicability of the solution. The defined marginal energy consumption coefficient directly describes the local slope relationship between providing reactive power support and the increased energy consumption cost under specific future operating conditions. This coefficient provides a key input for the subsequent multi-objective optimization decision model, enabling the system to accurately determine the economic inflection point for continuing operation, standby, or performing start-stop operations within a specific time period based on this quantitative relationship.

[0023] Preferably, in step (3), the multiple energy consumption assessment models include:

[0024] The start-stop transient loss assessment function includes at least mechanical impact loss, electromagnetic transient loss and insulation aging accelerated loss components.

[0025] The no-load excitation loss function includes at least the no-load iron loss component related to the square of the terminal voltage and the auxiliary power consumption component related to the cooling system load rate.

[0026] The load operation loss function includes at least a stator copper loss component, a stray loss component, and an efficiency degradation component dynamically corrected based on the marginal energy consumption coefficient.

[0027] Preferably, in step (3), generating the start / stop decision boundary vector includes:

[0028] By weighting and combining the various loss functions in the multiple energy consumption assessment models, a total objective function is constructed with the reactive power threshold corresponding to the start-stop decision boundary vector at each future time as the decision variable.

[0029] Solving the overall objective function yields the Pareto optimal solution set;

[0030] A selection strategy based on minimizing normalized deviation is adopted to select the optimal solution from the Pareto optimal solution set, forming the start-stop decision boundary vector.

[0031] Preferably, in step (3), the fuzzy logic processing includes:

[0032] Within a preset sliding time window, the real-time reactive power deficit is compared point by point with the start-stop decision boundary vector at the corresponding time, generating a preliminary decision sequence consisting of start flags and stop flags.

[0033] Using the trend statistics of continuous start-up and continuous shutdown in the preliminary sequence as fuzzy input, reasoning and defuzzification are performed through a predefined fuzzy rule base to obtain smooth decision values;

[0034] The smoothing decision value is compared with preset start-up and stop-up thresholds to generate the final start-up, maintain, or stop command.

[0035] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method of the present invention.

[0036] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the method described in the present invention.

[0037] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) Through unified modeling and multi-scale feature extraction of multi-source heterogeneous data, a highly consistent expression of equipment operating status, power grid demand changes and environmental impact is achieved; (2) By combining long short-term memory network and self-attention mechanism to establish dynamic energy consumption benchmark model, and integrating scheduling plan and load forecast data, the collaborative prediction of future reactive power demand and marginal energy consumption is achieved, breaking through the limitation of traditional methods that only predict demand and do not characterize energy consumption; (3) By constructing multiple energy consumption models including start-stop transient loss, no-load excitation loss and load operation loss, and introducing multi-objective optimization solution to form dynamic start-stop decision boundary, the start-stop control is transformed from a fixed threshold method to an adaptive optimization mechanism oriented towards future operating conditions; (4) By combining sliding time window comparison and fuzzy logic processing to achieve smooth output of control commands, the problem of frequent start-stop caused by short-term disturbances is effectively suppressed, and the stability and reliability of control are improved. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method of the present invention;

[0039] Figure 2 This is a schematic diagram of the computer device of the present invention;

[0040] Figure 3 is a start-stop control diagram of the present invention. Detailed Implementation

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0042] Example 1

[0043] As attached Figure 1As shown, this embodiment provides an automatic start-stop control method for synchronous condensers based on adaptive energy consumption optimization. This method addresses issues in existing synchronous condenser start-stop control systems, such as the lack of modeling capabilities for dynamic energy consumption characteristics, insufficient accuracy in multi-source data fusion and reactive power demand prediction, and poor decision-making economy due to the failure to consider comprehensive optimization of multiple energy consumption factors. The method includes the following steps:

[0044] Step 1: Obtain historical operating data of synchronous condensers, power grid reactive power demand sequence, equipment status monitoring data and environmental parameters, and perform data cleaning and time series alignment. Extract statistical features, frequency domain features and cross features to construct a standardized multidimensional feature matrix.

[0045] This step establishes an input representation system based on "learnable start-stop energy consumption, expressible demand fluctuations, and quantifiable equipment status." First, missing value imputation and outlier removal ensure the quality of single-channel data. Then, unified time-scale synchronization and 1-minute resampling establish comparability across channels at the same time. Next, multiple features reflecting operational levels, fluctuation patterns, and physical coupling relationships are extracted under a unified time coordinate. Finally, normalization forms a standardized multi-dimensional feature matrix that can be directly used for subsequent model training, thus providing high-quality, highly consistent input data for subsequent modeling. The specific implementation process is as follows:

[0046] 1. Data Acquisition and Cleaning

[0047] (1) Establish the original dataset. The original data includes 12 consecutive months of operating data obtained from the distributed control system of synchronous condensers, the power grid reactive power demand sequence collected by the synchronous phasor measurement unit, the stator winding insulation resistance sequence and cooling medium temperature sequence obtained by the equipment monitoring terminal, and the ambient temperature sequence and relative humidity sequence obtained from the meteorological station. The operating data includes the instantaneous value sequence of the three-phase voltage at the generator terminal, the instantaneous value sequence of the three-phase current of the stator, the reactive power regulation sequence, the stator winding hot spot temperature sequence, and the shaft vibration amplitude sequence sampled every 15 seconds. The original timestamps of the multi-source data are retained when entering the same dataset for unified alignment. All sequences are recorded as the original dataset, and the formula is:

[0048]

[0049] in, This represents the raw dataset that has not yet been cleaned; Indicates the first Physical quantities at time The original observations; Indicates data channel index ( ), used to distinguish different types of sampling variables; Indicates the index of the original sampling time ( (), used to distinguish observations of the same data channel at different sampling times; The total number of channels corresponding to synchronous condenser operation data, power grid reactive power demand, equipment status monitoring data, and environmental parameters; The total number of original sampling points during the corresponding observation period.

[0050] (2) Missing value handling: Missing time points in the original dataset are filled using cubic spline interpolation. The specific process is as follows:

[0051] For any data channel in the original dataset In the absence of time (located at adjacent valid sampling times) and (between), a piecewise cubic spline function is used for interpolation, and the interpolation formula is:

[0052]

[0053] in, Indicates missing moments The estimated value obtained through interpolation; , , and Indicates the first Each channel is in the interval The corresponding spline coefficient; subscript This indicates the interpolation interval number. The spline coefficients are jointly determined by constraints on the continuity of function values ​​at the interval endpoints, the continuity of the first derivative, and the continuity of the second derivative, to ensure that the interpolation curve does not have abrupt changes at the connection points of adjacent intervals. After interpolation is completed for all missing time points, the complete dataset is obtained.

[0054] (3) Outlier Handling: Considering that outliers in engineering implementation mainly correspond to non-physical spikes caused by sensor jumps, communication noise, or instantaneous interference in the field (such as instantaneous spikes in hot spot temperature measurement points, isolated high values ​​after vibration probes lose pulses, etc.), this invention adopts a "first identify, then remove, then repair" processing strategy, which can remove outlier data points from the source and use reasonable numerical values ​​to repair the sequence, thereby ensuring the purity of the sequences used to calculate the mean, variance, and spectrum, and avoiding contamination of the constructed feature matrix. The processing flow is as follows:

[0055] After imputing missing values, outliers exceeding three times the standard deviation in the original data were identified and removed using the Grubbs' test. For the... For each channel, first calculate its sample mean within the current observation window. and sample standard deviation Then, calculate the Grubbs discriminant statistic for each sampling point. The calculation formula is as follows:

[0056]

[0057] in, Indicates the first Each channel at time The degree of abnormal deviation. When Greater than the critical value at the preset significance level ( When the significance level can be determined in advance by looking up a table based on the sample length and significance level, then... If an outlier is identified, the outlier is removed from the current channel sequence, and then interpolation is performed using adjacent valid points in the same channel to repair it.

[0058] After completing missing value imputation and outlier removal, the cleaned dataset is obtained. .

[0059] 2. Timing alignment and resampling

[0060] After obtaining the cleaned dataset Subsequently, unified time-scale synchronization and resampling are performed to eliminate the impact of inconsistencies in sampling periods and clock references between different devices. The minute-level time sequence after unified time-scale synchronization is defined as follows: :

[0061]

[0062] in, Indicates the start time of the unified timeline; Indicates a uniform resampling period; Represents a minute-level index on a unified timeline; This represents the total number of moments after synchronization with a unified timescale. Considering that subsequent start / stop decision boundary vectors are output at the minute level, and that the 4-hour prediction window corresponds to exactly 240 time points, ensuring strict consistency with subsequent predictions and optimizations in terms of time granularity, thus avoiding feature input drift caused by time granularity mismatch, the resampling period in this embodiment is [value missing]. .

[0063] After defining a unified time scale for synchronization, the first All channels at the same time The resampled value For electrical quantity channels that change rapidly and have short sampling intervals, linear resampling is used, with the following formula:

[0064]

[0065] in, Located at adjacent original sampling time and between.

[0066] For state or environmental quantities that change relatively slowly, such as hot spot temperature, cooling medium temperature, insulation resistance, and relative humidity, the most recent hold method is adopted to keep the previous effective value unchanged between adjacent minute-level sampling.

[0067] After completing time-stamp synchronization and resampling, the aligned time-series dataset is obtained, represented as:

[0068]

[0069] in, This indicates that all channels have completed the minute-level alignment of the time-stamped dataset.

[0070] 3. Feature extraction and standardization

[0071] (1) Statistical feature extraction: The start-stop decision of the synchronous condenser depends not only on the instantaneous sampled value, but also on the stability and fluctuation characteristics of the operation over a continuous period. In order to take into account both minute-level control input and hour-level operating state distribution, the length is used as the starting point. A sliding statistical window is used to extract time-domain statistical features. The first... The sample set for the hourly window is :

[0072]

[0073] For any channel In the window Internal calculation mean ,variance and kurtosis The calculation formula is:

[0074]

[0075]

[0076]

[0077] in, Used to characterize the The first channel in the Average operating level within the hourly window; Used to characterize the volatility of the channel; This is used to characterize the sharpness of the waveform in the channel, thereby reflecting changes in load impact, frequency of control actions, or equipment operating stability. Indicates the first The first channel in the The standard deviation within the hourly window.

[0078] In the feature extraction process, this embodiment extracts and introduces such hourly statistical features, which can provide a comprehensive representation of the "distribution of operating status" of the equipment for the subsequent energy consumption model, rather than being limited to the description of the instantaneous state at a single point in time, thereby enhancing the model's overall assessment capability of the equipment's operating status.

[0079] (2) Frequency Domain Feature Extraction: In addition to time domain statistical features, this embodiment also performs frequency domain feature analysis to deeply explore the temporal patterns contained in reactive power regulation behavior. By performing a fast Fourier transform on the reactive power regulation sequence, its spectral energy distribution is obtained. This spectral energy distribution can effectively identify the characteristic differences of reactive power regulation at two levels: slow changes at low frequencies and frequent fluctuations at high frequencies. For automatic start-stop control, if there is a continuous energy accumulation of reactive power regulation in the higher frequency band, it usually indicates that the reactive power demand of the power grid has a high-frequency small-amplitude fluctuation. At this time, if the start-stop judgment is based only on the instantaneous value or fixed threshold, it is easy to cause ineffective frequent actions and increase equipment losses. Therefore, this embodiment incorporates frequency domain features into the feature matrix constructed later, aiming to enhance the model's ability to identify and adapt to the "rhythm of demand change" of the power grid and provide key temporal dynamic information for subsequent optimization decisions.

[0080] Defined in the Within the hourly window, the minute-level sequence corresponding to the reactive power regulation is as follows: subscript This represents the reactive power regulation channel. Performing a Fast Fourier Transform on this sequence yields the [number]th [sequence]. spectral coefficients of each frequency component :

[0081]

[0082] The spectral energy distribution is further calculated and expressed as:

[0083]

[0084] in, Indicates a local time index within the window; The imaginary unit, Indicates the first Within the first time window, the first The reactive power adjustment value corresponding to each sampling point. Indicates the discrete frequency component index. Indicates the number of sampling points within the time window; This represents the energy of the corresponding frequency component.

[0085] (3) Construction of cross features: Based on statistical features and frequency domain features, cross features are further constructed. The cross features are used to transform independently acquired physical quantities into combined quantities that can directly reflect the coupling state of the equipment, including aging sensitivity features composed of insulation resistance and winding temperature, mechanical load coupling features composed of vibration amplitude and current fluctuation, temperature rise gradient features composed of stator winding hot spot temperature and ambient temperature, and apparent power features composed of the effective value of terminal line voltage and the effective value of stator line current, etc. In this embodiment, temperature rise gradient features and apparent power features are used.

[0086] (31) Constructing a temperature rise gradient feature sequence. Since the heat dissipation pressure corresponding to the same hot spot temperature is different under different ambient temperatures, the hot spot temperature itself cannot accurately describe this difference. Therefore, a temperature rise gradient feature sequence is constructed. Used to characterize the degree of heat accumulation and cooling load level:

[0087]

[0088] in, Indicates time The stator winding hot spot temperature; Indicates time The ambient temperature.

[0089] (32) Constructing the apparent power characteristic sequence. Since combining the two independent measurements of voltage and current into a comprehensive index that can reflect the electromagnetic load level makes it easier for subsequent models to learn the theoretical energy consumption baseline variation law under different operating points, this embodiment constructs the apparent power characteristic sequence. :

[0090]

[0091] in, This indicates the effective value of the terminal line voltage. This represents the effective value of the stator line current.

[0092] (4) Feature standardization: All features are concatenated into a feature vector in a uniform order. ,in , The total feature dimension, This represents the feature dimension index. To eliminate the impact of inconsistent feature dimensions on subsequent model training, min-max normalization is performed on each feature dimension to obtain standardized features. :

[0093]

[0094] in, Indicates the first dimensional features at time The original value; and These represent the minimum and maximum values ​​of this feature within the statistical range of the training samples, respectively. A very small positive number is set to prevent the denominator from being zero (in engineering implementation, Desirable Or other small constants that will not affect the normalization result.

[0095] Through the above processing, all features are compressed into a uniform numerical range. This avoids high-dimensional features dominating the loss function during model training and facilitates the establishment of stable gradient update relationships for features from different sources by the subsequent Long Short-Term Memory network. Finally, a standardized multidimensional feature matrix is ​​constructed. :

[0096]

[0097] in, The rows correspond to the unified minute-level time steps Columns correspond to feature dimensions .

[0098] Through the aforementioned data preprocessing and multi-scale feature extraction mechanism, this invention achieves highly consistent feature representation of multi-source information such as synchronous condenser operating status, power grid reactive power demand, and environmental parameters. Specifically, addressing the problem of insufficient representation caused by relying on single electrical quantities or simple statistical values ​​in existing technologies, this embodiment performs data cleaning, anomaly removal, and unified time-scale alignment, and jointly constructs time-domain statistical features, frequency-domain energy features, and physical cross-features. This transforms the originally discrete and noisy raw data into a structured feature matrix with both clear physical meaning and temporal consistency, thereby effectively improving the ability of the input data to represent real operating conditions. Furthermore, by introducing cross-features such as temperature rise gradient and apparent power, coupled representation of key internal physical quantities such as equipment thermal state and electromagnetic load state is achieved, thus solving the deficiency of traditional methods in insufficient characterization of multi-dimensional state correlations within equipment and providing a reliable data foundation for subsequent high-precision modeling.

[0099] Step 2: Input the standardized multidimensional feature matrix into the long short-term memory network and introduce a self-attention mechanism to generate a dynamic energy consumption benchmark model for the equipment. Then, fuse it with the power grid dispatch plan data and load forecast data, and input it into a one-dimensional convolutional neural network to obtain the future reactive power demand forecast value and marginal energy consumption coefficient.

[0100] This step involves constructing a dynamic energy consumption baseline model for equipment using a standardized multidimensional feature matrix, obtaining a theoretical energy consumption baseline sequence for the next 240 minutes to characterize the intrinsic energy consumption trend of the equipment under the current operating conditions. This theoretical energy consumption baseline sequence, along with future scheduling plans and load forecasts, is then input into a time-series convolutional network to obtain a future reactive power demand forecast sequence and a marginal energy consumption coefficient sequence. The future reactive power demand forecast sequence serves as the input to external demand constraints, while the marginal energy consumption coefficient sequence serves as a key input to the operational energy consumption sensitivity in multi-objective optimization, thus forming a continuous technical chain of "historical operating condition characterization – energy consumption baseline learning – future demand coupled forecasting." The specific implementation process is as follows:

[0101] 1. Construct a dynamic energy consumption benchmark model for equipment.

[0102] (1) Model Input: Define the length of the historical input window used for theoretical energy consumption baseline modeling as In this embodiment, This involves using 240 consecutive minutes of samples prior to the current decision-making moment as a set of historical inputs. By selecting 240 consecutive minutes of features as inputs, the system covers the complete short-cycle operating trajectory of the synchronous condenser from hot standby, no-load, load regulation to deep leading / lagging phase conditions. At the same time, it aligns the subsequent energy consumption baseline output with the prediction window of the next 240 minutes point by point, facilitating direct integration with scheduling plans and load forecasting data.

[0103] For the current decision anchor Its input sequence is defined as :

[0104]

[0105] Among them, the current decision anchor point As the endpoint, To unify the index of the current moment on the timeline, For historical input sequences; Indicates the current decision anchor point Output 3D standardized feature column vector; Indicates the length of the history window.

[0106] (2) Network structure: A long short-term memory network (LSTM) is constructed, with 128 memory units in its hidden layer. Through the gating mechanism, the network can retain time step information that contributes to the evolution of energy consumption within the historical window, suppress random disturbances and short-term fluctuations that are unrelated to energy consumption, thereby adapting to the slow heating process and fast response process that exist simultaneously in the operation of the synchronous condenser.

[0107] Input sequence Before inputting the Long Short-Term Memory (LSTM) network, it is expanded into sub-vectors in chronological order. :

[0108]

[0109] in, Indicates the number within the history window Input features at each time step, This represents the index of the time step within the historical time window. .

[0110] For the sequence in the th At each time step, the input gate, forget gate, candidate state, and output gate operations are performed. The state update can be represented as:

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117] in, , and These represent the input gate, forget gate, and output gate at the [number]th ... Gated vectors at each time step; Represents the candidate state vector; Represents the current memory state vector; This represents the current hidden output vector. Indicates the current decision-making moment The next historical sequence The hidden state vector corresponding to each time step; , , , This is the weight matrix input to each gate; , , , This is the recursive weight matrix from the previous hidden state to each gate; , , , It is the bias vector; This represents the Sigmoid activation function; Represents the hyperbolic tangent function; symbol This indicates element-wise multiplication.

[0118] (3) Introduction of self-attention mechanism: Considering that the terminal output of the Long Short-Term Memory network can extract certain temporal correlations, it lacks the ability to explicitly distinguish the importance of historical information, making it difficult to automatically focus on the key time period of the future energy consumption baseline within the historical window. Meanwhile, the theoretical energy consumption baseline of the synchronous condenser is not only related to the current operating conditions, but also to the heat accumulation, insulation attenuation overpotential, and load adjustment frequency over a previous period, and the contributions of these influences on the time axis are not uniform. Therefore, this invention introduces a self-attention mechanism into the Long Short-Term Memory network framework to assign weights to all hidden outputs within the historical window. By assigning weights through the self-attention mechanism, the model can automatically focus on moments more relevant to the future energy consumption baseline (such as the cooling transition phase after long-term high-temperature operation, periods of frequent reactive power adjustment, or periods of disturbance accumulation before the start-stop boundary). The specific process is as follows:

[0119] (31) Calculate attention score. Attention score at each time step :

[0120]

[0121] in, This represents a trainable attention query vector. Indicates the transpose operation; and These are the mapping weight matrix and bias vector of the attention layer, respectively.

[0122] (32) Calculate the attention weights. The attention weights are obtained by Softmax normalization. :

[0123]

[0124] in, Indicates the in-window time index used during normalized summation.

[0125] (33) Weighted summation. The context vector is obtained by weighted summation of all hidden outputs. :

[0126]

[0127] (34) Obtain the enhanced feature vector. The enhanced feature vector is formed by concatenating the context vector with the hidden output of the last time step. :

[0128]

[0129] Among them, symbols This indicates vector concatenation.

[0130] (4) Model training and output

[0131] (41) Calculate the theoretical energy consumption baseline output. Enhance the eigenvector. This is mapped to a minute-by-minute theoretical energy consumption baseline sequence for the next 4 hours. The number of future prediction steps is defined as... The theoretical energy consumption baseline output is:

[0132]

[0133] in, Indicates the current decision anchor point The theoretical energy consumption baseline sequence for the next 240 minutes. To output the mapping matrix; To output the bias vector, This is the theoretical energy consumption baseline value predicted by the model.

[0134] (42) Training is performed using the mean squared error loss function. The future energy consumption sequence obtained statistically or by measurement under actual operating conditions is used as the supervision label, and the mean squared error loss function is used for parameter optimization. Among them, the future energy consumption sequence... for:

[0135]

[0136] Energy consumption baseline modeling loss for:

[0137]

[0138] in, Indicates the future number The actual energy consumption label value per minute.

[0139] (43) Establish a dynamic energy consumption benchmark model for the equipment. Iteratively update all parameters using the BPTT backpropagation algorithm and monitor the loss convergence on the validation set. When the... The training cycle and the first When the validation set loss for each training epoch meets a certain condition, and this condition is met continuously for 10 training epochs, the energy consumption baseline model is considered to have converged stably. Subsequently, the network parameters are fixed to form a dynamic energy consumption baseline model for the device. Using the above convergence criterion avoids premature termination due to slight fluctuations in loss during the later stages of training, and also avoids the model continuing to overfit on noisy samples. The condition is:

[0140]

[0141] in, This represents the mean squared error loss for energy consumption baseline modeling, where E indicates that this loss function is the energy consumption baseline modeling loss. This indicates the validation set, meaning the loss value is calculated on the validation set, not the training set. It's used to monitor the model's generalization ability and prevent overfitting. For training cycle indexing, This indicates the previous training cycle.

[0142] Once trained and fixed, the equipment dynamic energy consumption benchmark model can be used to determine the energy consumption of a given historical standardized feature sequence. Output the theoretical energy consumption baseline sequence for the next 4 hours. This sequence reflects the intrinsic energy consumption evolution trend of the synchronous condenser without considering external scheduling constraints, under the current operating state and the continuation of the historical state.

[0143] 2. Multi-source data fusion and prediction

[0144] (1) Data fusion: Obtain dispatch plan data for the next 4 hours from the power grid dispatch center, including the bus voltage planned value sequence. And the planned reactive power output sequence Obtain the minute-by-minute load forecast sequence for the next 4 hours from the load forecasting system. The theoretical energy consumption baseline value sequence By concatenating the above three data sources according to their input channels, a multi-source heterogeneous data fusion tensor is constructed. (Dimensions are 4×240):

[0145]

[0146] in, Indicates the current decision-making moment Next The first feature channel in the future The fused data values ​​corresponding to each time step For feature channel index, .

[0147] The multi-source heterogeneous data fusion tensor This provides a structured, unified input for subsequent temporal convolutional networks, enabling them to capture the temporal coupling relationship between "equipment energy consumption, reactive power planning, voltage planning, and load changes" within a unified tensor structure. In this fused tensor, by placing the theoretical energy consumption baseline in the first channel and arranging scheduling plan data and load prediction sequences in subsequent channels, the temporal convolutional network can simultaneously perceive the coordination between intrinsic equipment energy consumption changes and external scheduling drivers, avoiding one-sided predictions based on single-dimensional information. The four channels of the final fused tensor correspond one-to-one with 240 future minute-level time points, ensuring that any column in the fused tensor represents the complete decision-making environment for the same future minute.

[0148] (2) Temporal Convolutional Network Prediction: Construct a temporal convolutional network with 3 convolutional layers, and fuse tensors The input is used to extract local trends, mutation patterns, and cross-channel coupling relationships in the future time domain. Each layer of the temporal convolutional network has 64 one-dimensional causal convolutional kernels of size 3, and the ReLU activation function is used. The nth... layer( The input feature map for (=1,2,3) is: The output feature map is And the initial input .

[0149] Then the first The first in the layer Each output feature map at time position eigenvalues ​​at The calculation is as follows:

[0150]

[0151] in, Indicates the index of the output feature map in the corresponding layer; Indicates the first The number of input channels in the layer, where d represents the input channel index; This represents the offset of the causal convolution kernel on the time axis. Since the kernel size is 3, therefore... , Indicates the first Layer Each convolutional kernel corresponds to one input channel. In offset Convolution weights at the point, Indicates the first The first in the layer Each input channel at time position Eigenvalues ​​at; Indicates the corresponding bias term; This represents the ReLU activation function.

[0152] In implementation, when When the boundary is crossed, the feature values ​​are filled with zero. The use of causal convolution ensures that the model can predict the future... At the minute level, relying only on information from the current and historical moments without revealing information from subsequent moments, it meets the causality requirements of time series forecasting.

[0153] After each convolutional layer, a max-pooling operation with a stride of 2 is performed to compress the temporal dimension and enhance the main trend features. Convolutional output The corresponding max pooling results for:

[0154]

[0155] in, This is the time index after pooling. Pooling output of layers As the next level (i.e., the first) +1 layer) input .

[0156] Through the three layers of causal convolution and pooling operations described above, the network can extract high-order temporal features at different time scales from the fused tensor: the first layer mainly captures local response relationships within adjacent minutes (such as the impact of planned reactive power changes on the short-term energy consumption baseline); the second layer mainly captures trend segments on the order of ten minutes (such as the coordinated pattern of voltage planning adjustments and load increases / decreases); and the third layer integrates the cascading relationships of cross-channel mesoscale correlations (such as the cascading relationship of "load increase - planned reactive power increase - energy consumption baseline increase"). This multi-layer structure enables the model to shift from relying on numerical comparisons at a single moment to being able to identify temporal patterns that determine changes in start-stop boundaries.

[0157] After completing the final pooling layer, the output feature map is flattened to obtain the feature vector. .

[0158] (3) Dual output head: converts the feature vector The input is a fully connected layer with dual output heads (containing 128 neurons), which simultaneously outputs a sequence of reactive power demand predictions for the next 240 minutes. and marginal energy consumption coefficient sequence , respectively represented as:

[0159]

[0160]

[0161] in, and These represent the weight matrices for the reactive power demand output head and the marginal energy consumption coefficient output head, respectively. and These are the corresponding bias vectors.

[0162] also, It is a comprehensive prediction of future actual reactive power demand based on the combined effects of theoretical energy consumption baseline, bus voltage planning, planned reactive power output, and load forecasting, rather than directly using the planned reactive power value from the dispatch plan. Therefore, it can be directly used as the demand input for start-stop boundary optimization. Marginal energy consumption coefficient Defined as the rate of change of energy consumption corresponding to a unit increment of reactive power command.

[0163] To make the marginal energy consumption coefficient trainable and quantifiable, a marginal energy consumption label is constructed during the model training phase. A fixed reactive power perturbation is set as... (For example: take) Mvar). For the future nth training sample. The plan for minutes of unproductive output Applying a unit perturbation, we obtain And maintain the multi-source heterogeneous data fusion tensor The data for the remaining channels remains unchanged. This perturbed sample is then re-input into the trained, fixed-parameter device dynamic energy consumption baseline model to obtain the predicted energy consumption baseline value. Based on this, a marginal energy consumption label is defined. for:

[0164]

[0165] in, Indicates the future of the original sample The actual energy consumption label value corresponding to each minute. This represents the theoretical energy consumption baseline prediction after reactive power perturbation.

[0166] To simultaneously train both the future reactive power demand forecast and the marginal energy consumption coefficient as output heads, the temporal convolutional network employs a joint loss function. Conduct training:

[0167]

[0168] in, This represents the joint training loss of a temporal convolutional network. Indicates the future number Real label of reactive power demand in minutes; This indicates the corresponding marginal energy consumption label; and These represent the weighting coefficients for the two types of loss terms, used to balance the importance of accuracy in future reactive power demand and accuracy in marginal energy consumption sensitivity. In engineering implementation, different optimization emphases can be achieved by configuring the relevant weights: for higher prediction accuracy, these weights can be appropriately increased. If more attention is paid to the quality of economic optimization at the subsequent start-stop boundaries, then the level can be appropriately increased. .

[0169] After training, fix the convolutional network parameters and at any current decision anchor point Input fusion tensor This allows for the simultaneous output of future reactive power demand forecast sequences. With marginal energy consumption coefficient series .

[0170] By constructing a coupled modeling framework of energy consumption baseline modeling and multi-source fusion prediction, this invention achieves the coordinated prediction of future reactive power demand and marginal energy consumption, thus overcoming the limitation of the separation between demand forecasting and energy consumption analysis in existing technologies. This invention utilizes a long short-term memory network combined with a self-attention mechanism to extract key time-series information with the greatest influence on energy consumption changes from historical features, establishing a dynamic energy consumption baseline model. This allows for accurate characterization of the energy consumption evolution trend of equipment under different operating conditions. Furthermore, this energy consumption baseline is fused with power grid dispatching plans and load forecasting data, and the coupling relationship between multi-source data is extracted through a temporal convolutional network, simultaneously outputting the predicted value of future reactive power demand and the marginal energy consumption coefficient. Through this design, the model can simultaneously complete reactive power demand prediction and provide reactive power cost assessment, thus providing a dual-dimensional forward-looking basis for subsequent start-up and shutdown decisions, integrating demand and cost, and improving the economic basis and prediction accuracy of the decision-making process.

[0171] Step 3: Based on the predicted future reactive power demand and marginal energy consumption coefficient, construct multiple energy consumption models and perform multi-objective optimization to generate start-stop decision boundary vectors. At the same time, compare the start-stop decision boundary vectors with the real-time reactive power deficit through a sliding window and output the start-stop control command of the synchronous condenser after fuzzy logic processing.

[0172] This step combines multiple energy consumption models with multi-objective optimization methods, supplemented by a sliding time window and fuzzy logic decision smoothing mechanism, to achieve the transformation of synchronous condenser start-stop control from fixed threshold to dynamic adaptive optimization. The process mainly includes: First, constructing multiple energy consumption models encompassing start-stop transient losses, no-load excitation losses, and load operating losses; introducing marginal energy consumption coefficients to dynamically correct operating efficiency; and generating a time-varying start-stop decision boundary vector through multi-objective optimization. Second, performing trend analysis on real-time reactive power deficit using a sliding time window; and using fuzzy logic to smooth the initial decision, effectively suppressing malfunctions and frequent start-stop problems caused by short-term fluctuations. Finally, outputting stable and executable start-stop control commands, thereby ensuring the reactive power support capacity of the power grid while improving equipment operating economy and extending service life. The specific implementation process is as follows:

[0173] 1. Constructing and optimizing various energy consumption models.

[0174] (1) For the current decision anchor point The next chapter Within minutes, a start-stop transient loss evaluation function consisting of mechanical impact loss components, electromagnetic transient loss components, and insulation aging acceleration components is constructed. :

[0175]

[0176] in, Indicates the future number The mechanical shock loss caused by performing start-stop operations every minute. This represents the corresponding electromagnetic transient loss. This represents the reduced lifespan due to accelerated insulation aging. The three components of loss are defined as follows:

[0177]

[0178]

[0179]

[0180] in, It represents the mechanical loss coefficient, which is used to characterize the cost reduction factor of shaft impact, coupling stress fluctuation and additional bearing wear caused by a single start-stop; This indicates that the current optimization scheme will be used in the future. The cumulative number of start-stop operations per minute; It represents the electrical loss coefficient, used to characterize the effect of inrush current on the electromagnetic stress of the winding and excitation system; Indicates the future number The expected inrush current amplitude during the minute-long start-stop process; Indicates the insulation aging cost coefficient; Indicates the future number Predicted winding hot spot temperature values ​​for each minute; Indicates the future number The insulation aging factor is calculated in minutes. To accurately quantify the differential impact of start-up and shutdown operations on the insulation life of equipment at different winding temperatures, a winding temperature-related correction term is directly introduced into the insulation aging item.

[0181] Considering that insulation degradation is usually not a linear process, but is related to historical heating levels, moisture levels, and the accumulation of electrical stress, an exponential decay model can better characterize its initial gradual decline followed by accelerated degradation. Therefore, the insulation aging factor... The formula is based on an exponential model of insulation resistance decay over time:

[0182]

[0183] in, Indicates the current decision anchor point The insulation resistance measured at all times, Indicates the rated reference insulation resistance. Indicates the insulation attenuation coefficient. This represents the predicted insulation resistance at the f-th minute in the future, assuming the current insulation degradation trend continues. This indicates the corresponding insulation aging factor; s represents a uniform time step in minutes. The lower the value, The larger the value, the more vulnerable the equipment insulation is. In this case, the impact of the same start-stop action on the lifespan should be amplified and taken into account.

[0184] (2) Regarding the current decision anchor point Next time Minutes are used to construct the no-load excitation loss function, which is composed of the no-load iron loss component and the auxiliary power consumption component. :

[0185]

[0186] in, Indicates the no-load iron loss component. This represents the power consumption component of the auxiliary system.

[0187] To accurately reflect the correlation between the iron loss of the synchronous condenser and the random terminal voltage level in the excitation standby state, a relationship proportional to the square of the terminal voltage is introduced into the no-load iron loss component. The auxiliary system power consumption component encompasses the basic energy consumption of equipment necessary to maintain the equipment's standby state, such as the lubricating oil pump, cooling fan, heat exchanger, and control power supply; its actual power consumption dynamically changes with the cooling load rate. The no-load iron loss component and the auxiliary system power consumption component are expressed as follows:

[0188]

[0189]

[0190] in, Indicates the rated no-load iron loss; Indicates the future number The planned value of bus voltage or equivalent value of terminal voltage for each minute; Indicates the rated reference voltage; Indicates the rated power consumption of the auxiliary system; This indicates the load rate of the cooling system.

[0191] Under constant internal thermal conditions (such as winding heating), higher ambient temperatures or larger temperature rises in the cooling medium will lead to an increase in the operating pressure index, which in turn significantly increases the power consumption required for the cooling system to maintain standby operation. This change directly affects the economic boundary between hot standby and complete shutdown. Therefore, defining the cooling system load rate is crucial. The ambient temperature and the state of the cooling medium are quantified together as the operating pressure index of the auxiliary system, quantifying the impact of environmental conditions and the internal thermal state of the equipment on the operating intensity of the auxiliary system. The formula is as follows:

[0192]

[0193] in, Indicates the future number Minute-by-minute ambient temperature forecast Indicates the future number Predicted cooling medium temperature for minutes This is the reference temperature difference under rated operating conditions. If If, then it can be truncated to 0; if If so, it can be truncated to 1 or processed according to the maximum allowable limit on site.

[0194] (3) Regarding the current decision anchor point Next time Minutes were used to construct a load operation loss function consisting of stator copper loss components, stray iron loss components, and efficiency degradation components. :

[0195]

[0196] in, Indicates the amount of copper loss. Indicates stray iron loss. This represents the additional loss component resulting from efficiency degradation. The three parts are represented as follows:

[0197]

[0198]

[0199]

[0200] in, Indicates the stator equivalent resistance; Indicates the future number Predicted stator current value per minute; Indicates the stray iron loss coefficient; Indicates the future number The equivalent electric angular frequency or magnetic field angular frequency per minute; Represents the equivalent magnetic flux density at the corresponding moment; Indicates the future number The non-functional equivalent power term of the device output per minute; Indicates rated efficiency; Indicates the future number The predicted operating efficiency value for each minute. The efficiency prediction value is not directly treated as a constant, but is corrected using the output marginal energy consumption coefficient, expressed as:

[0201]

[0202] in, A coefficient representing the mapping from marginal energy consumption coefficient to the degree of efficiency degradation; Indicates the future number of outputs Minute marginal energy consumption coefficient.

[0203] The marginal energy consumption coefficient is transformed into a corresponding efficiency degradation component and incorporated into the load operation loss function. The purpose is to establish a connection between data-driven energy consumption sensitivity and analytical physical loss model. This design ensures that the energy consumption assessment used for subsequent optimization decisions is not only based on empirical formulas but also explicitly incorporates the prediction results output from step 2.

[0204] (4) Multi-objective optimization generates decision boundaries: for the future... Minutes, define the decision variable as This represents the reactive power threshold corresponding to the post-stop decision boundary vector at that moment. The start-stop transient loss assessment function, no-load excitation loss function, and load operation loss function are combined according to their weights to obtain the overall objective function:

[0205]

[0206] in, Indicates the current decision anchor point The overall optimization objective is as follows; , and These represent the weighting coefficients for start-stop transient losses, no-load excitation losses, and load operating losses, respectively, satisfying... By traversing or evolutionarily searching for feasible solutions under different weight combinations, the Pareto optimal solution set can be obtained. .

[0207] Candidate solutions in the Pareto front are defined as follows: :

[0208]

[0209] in Index for candidate solutions Indicates the first In subsequent decisions or iterations, at the current decision moment Corresponding to the future The start / stop decision boundary value at each time step.

[0210] Select the candidate boundary vector with the best overall compromise from the Pareto optimal solution set. (Achieved through a strategy that minimizes normalized deviation) The final start / stop decision boundary vector is calculated using the following formula:

[0211]

[0212] in, Indicates candidate solutions In the Cumulative cost value on class targets These correspond to start-stop transient losses, no-load excitation losses, and load operating losses, respectively. This represents the index of the objective function in multi-objective optimization, used to distinguish different types of optimization objectives (such as start-stop loss, no-load loss, and running loss). and These represent the Pareto fronts of the th... Minimum and maximum values ​​of the target class; To prevent extremely small positive numbers with a denominator of zero.

[0213] When determining the final start-stop decision boundary vector from the Pareto optimal solution set, a selection strategy based on minimizing normalized deviation is adopted. The aim is to select the solution with the best overall balance among various energy consumption optimization objectives from a set of non-dominated candidate solutions, thereby avoiding excessive bias in the decision result towards a single loss objective. The final determined start-stop decision boundary vector... Represented as:

[0214]

[0215] in Indicates the future number The optimal reactive power threshold point corresponding to each minute. This vector and the output future reactive power demand prediction sequence. They correspond one-to-one in the time dimension and serve as the benchmark boundary for real-time control criteria.

[0216] 2. Decision smoothing and instruction output (as attached) Figure 3 (As shown)

[0217] (1) Sliding window comparison

[0218] Deployment length is A sliding time window filter. At the current decision anchor point. The above extracts the most recent 60 minutes of real-time reactive power deficit to form a window sequence. :

[0219]

[0220] At the same time, from the start / stop decision boundary vector In the middle, select the anchor point that is most relevant to the current decision. The start / stop decision threshold corresponding to a given time is determined through a predicted time mapping relationship, i.e., taken as... The threshold point corresponding to the first minute in the future is used as the current real-time control reference value. Each value in the window sequence... Threshold at the corresponding time in the start / stop decision boundary vector Comparison, generating an initial decision sequence of length 60. :

[0221]

[0222] in, For each value in the window sequence The initial comparison indicator is constructed as follows: when the real-time reactive power deficit is higher than the current threshold, a startup trend is considered to exist; otherwise, no startup trend is considered to exist. The calculation formula is:

[0223]

[0224] in, This is the internal index of the sliding window.

[0225] (2) Fuzzy logic processing

[0226] Preliminary decision sequence Although it has already reflected the tendency of short-term start-stop, since it comes directly from minute-by-minute threshold comparison, it may still contain glitches caused by measurement noise, prediction bias or instantaneous reactive power disturbance. Therefore, this invention constructs a Mamdani-type fuzzy inference system for further fuzzy logic smoothing.

[0227] (21) Constructing fuzzy input. From the initial decision sequence Extract the number of consecutive start flags and the number of consecutive shutdown flags As a fuzzy input variable. This represents the number of consecutive "1"s counted backward from the end of the current window, used to characterize the persistence of the initiation trend; This represents the number of consecutive "0"s counted backward from the end of the current window, used to characterize the persistence of the shutdown trend. It uses a combination of triangles and trapezoids, respectively... and Define three fuzzy sets: low, medium, and high, and define their membership functions as follows: , , as well as , This embodiment uses the "high" membership degree of the number of consecutive starts and consecutive stops as an example:

[0228] Define the "high" membership degree of the number of consecutive starts as :

[0229]

[0230] Define the "high" membership degree of the number of consecutive shutdowns as :

[0231]

[0232] in, , , , This is a boundary parameter for a fuzzy set, and its value can be determined based on statistical analysis of historical start-stop data from the field.

[0233] When constructing the input of the fuzzy inference system, the number of consecutive start or stop flags within the sliding window is used as the fuzzy inference input. This can reflect the continuous trend of reactive power deficit exceeding the threshold and effectively distinguish between real control needs and short-term disturbances.

[0234] (22) Establish a rule base. After constructing the fuzzy input, a rule base is further established. This embodiment includes at least the following rules:

[0235] when For high and When the value is low, the output "startup" confidence level is high;

[0236] when for low and When the value is high, the output "shutdown" confidence level is high;

[0237] When both are in the middle, output "Maintain" state.

[0238] (23) Defuzzification. Defuzzification is performed on all rules in the rule base using the centroid method, and the smoothed decision value is calculated. :

[0239]

[0240] in, For the total number of rules, For rule indexing; Indicates the current decision anchor point Next The trigger strength of the rule; This represents the output center value of the corresponding rule.

[0241] 3. Command Generation and Issuance: In the design and configuration of the fuzzy inference system, to ensure the smooth decision values ​​output by defuzzification... It can be directly used for control decisions, setting the output center value of start-up rules to a positive value, the output center value of stop-up rules to a negative value, and the output center value of maintenance rules to zero. Based on this setting, the algebraic sign of the obtained value reflects the decision direction (positive for start-up, negative for stop-up), while the magnitude of its absolute value represents the confidence level of the decision.

[0242] Smooth decision values The code is compared with preset positive and negative decision thresholds to generate the final start / stop control command code. :

[0243]

[0244] in, Indicates the current decision anchor point The control command encoding below; The threshold for positive initiation decision; This is the threshold for a negative shutdown decision, and it usually satisfies... .when When indicates the generation of startup instructions; when "When" indicates maintaining the current state; "when" indicates maintaining the current state. The time indicates that a stop command has been generated.

[0245] Then, the instruction is encoded. The code is encapsulated according to the Modbus RTU protocol stack and sent to the synchronous condenser excitation regulator PLC. This protocol was chosen because of its direct compatibility with field PLCs and auxiliary control units, which facilitates the stable mapping of upper-level optimization decision commands to the actual equipment control layer. The PLC controls the coordinated actions of the static inverter, excitation system, lubrication system, and cooling system according to a preset start-stop sequence logic, thereby achieving automatic start-stop control.

[0246] Example 2

[0247] Corresponding to the method described in Embodiment 1, the present invention also provides an automatic start-stop control system for synchronous condensers based on energy consumption adaptive optimization, the system comprising:

[0248] The data processing module is used to perform the operation described in step 1 of embodiment 1, namely, to acquire historical operating data of synchronous condensers, power grid reactive power demand sequence, equipment status monitoring data and environmental parameters, and to perform data cleaning, time series alignment and feature extraction to construct a standardized multidimensional feature matrix.

[0249] The energy consumption calculation module is used to perform the operation described in step 2 of embodiment 1, that is, to construct a dynamic energy consumption benchmark model of the equipment based on the standardized multidimensional feature matrix, and to integrate scheduling and load forecasting data, and to predict future reactive power demand and marginal energy consumption coefficient through a time-series convolutional network.

[0250] The instruction output module is used to perform the operation described in step 3 of embodiment 1, that is, to construct multiple energy consumption models based on the predicted values ​​and perform multi-objective optimization to generate start-stop decision boundaries, and then combine sliding window comparison and fuzzy logic processing to output start-stop control instructions for the synchronous condenser.

[0251] Each module can be implemented based on an independent hardware unit, or it can be integrated into the same computing device and work together through software or firmware.

[0252] Example 3

[0253] The present invention also provides a computer device. The device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the automatic start / stop control method for a synchronous condenser as described in Embodiment 1.

[0254] As attached Figure 2As shown, the computer equipment can be a server, industrial control computer, or embedded edge computing gateway deployed in a power plant or dispatch center. It interacts with the distributed control system of synchronous condensers, synchronous phasor measurement units, and equipment status monitoring terminals via the IEC 61850 MMS communication interface and Modbus TCP / RTU dual-redundant interface. Simultaneously, it interacts with the power grid dispatch center's energy management system, automatic voltage control system, and ultra-short-term load forecasting system through a dedicated dispatch data network channel, and achieves real-time synchronization of environmental parameters with environmental meteorological monitoring stations via the OPC UA interface.

[0255] Example 4

[0256] The present invention also provides a computer-readable storage medium, such as a disk, optical disk, solid-state drive, or read-only memory, on which a computer program is stored. When the computer program is loaded and executed by a processor, it can realize the automatic start-stop control method for the synchronous condenser as described in Embodiment 1.

Claims

1. A method for automatic start / stop control of a synchronous condenser based on adaptive energy consumption optimization, characterized in that, Includes the following steps: (1) Obtain historical operating data of synchronous condensers, power grid reactive power demand sequence, equipment status monitoring data and environmental parameters and perform preprocessing, extract statistical features, frequency domain features and cross features that reflect the internal state coupling of equipment, and construct a standardized multidimensional feature matrix; (2) Based on the standardized multidimensional feature matrix, construct a dynamic energy consumption benchmark model for outputting the future theoretical energy consumption baseline; fuse the future theoretical energy consumption baseline with the power grid dispatch plan data and load forecast data to generate a multi-source heterogeneous data fusion tensor; The multi-source heterogeneous data is fused into a tensor and input into a temporal convolutional network to obtain a sequence of predicted reactive power demand values ​​and a sequence of marginal energy consumption coefficients for future periods; the equipment dynamic energy consumption benchmark model is implemented using a long short-term memory network combined with a self-attention mechanism. During the training phase of the device's dynamic energy consumption benchmark model, marginal energy consumption labels for training the marginal energy consumption coefficients are constructed in the following manner: Apply a preset fixed reactive power perturbation to the planned reactive power output values ​​in the sample; Input the sample containing the planned reactive power output value after the perturbation into the trained dynamic energy consumption benchmark model of the equipment, and calculate the theoretical energy consumption baseline prediction value after the perturbation. The marginal energy consumption label is constructed using the following formula: in, Labeling as marginal energy consumption; This is a preset fixed reactive power perturbation; This serves as the theoretical energy consumption baseline after the disturbance. Indicates the future of the original sample The actual energy consumption label value corresponding to each minute; (3) Based on the reactive power demand prediction value sequence and the marginal energy consumption coefficient sequence, construct a multi-type energy consumption assessment model including start-stop transient loss, no-load excitation loss and load operation loss, and generate start-stop decision boundary vector corresponding to future time points through multi-objective optimization solution; compare the real-time monitored reactive power deficit with the start-stop decision boundary vector within the sliding time window, and after fuzzy logic processing, finally output the synchronous condenser start-stop control command; The various energy consumption assessment models include: The start-stop transient loss assessment function includes at least mechanical impact loss, electromagnetic transient loss and insulation aging accelerated loss components. The no-load excitation loss function includes at least the no-load iron loss component related to the square of the terminal voltage and the auxiliary power consumption component related to the cooling system load rate. The load operation loss function includes at least a stator copper loss component, a stray loss component, and an efficiency degradation component dynamically corrected based on the marginal energy consumption coefficient.

2. The method according to claim 1, characterized in that, In step (1), the preprocessing includes missing value imputation, outlier removal and unified time scale alignment; the missing value imputation is performed using cubic spline interpolation; the outlier removal uses the Grubbs test to identify and remove outlier data points that exceed 3 times the standard deviation.

3. The method according to claim 1, characterized in that, In step (1), the cross-features include aging sensitivity features composed of insulation resistance and winding temperature, mechanical load coupling features composed of vibration amplitude and current fluctuation, temperature rise gradient features composed of stator winding hot spot temperature and ambient temperature, and apparent power features composed of the effective value of terminal line voltage and the effective value of stator line current.

4. The method according to claim 1, characterized in that, In step (2), the construction of the device dynamic energy consumption benchmark model includes the following steps: Based on the standardized multidimensional feature matrix, the output vector of the long short-term memory network at each time step is obtained; Calculate the self-attention weights and sum the output vectors at each time step to obtain the context vector; The context vector is concatenated with the final time-end output of the long short-term memory network to form a comprehensive feature vector; The integrated feature vector is mapped to the future theoretical energy consumption baseline.

5. The method according to claim 1, characterized in that, In step (3), generating the start / stop decision boundary vector includes: By weighting and combining the various loss functions in the multiple energy consumption assessment models, a total objective function is constructed with the reactive power threshold corresponding to the start-stop decision boundary vector at each future time as the decision variable. Solving the overall objective function yields the Pareto optimal solution set; A selection strategy based on minimizing normalized deviation is adopted to select the optimal solution from the Pareto optimal solution set, forming the start-stop decision boundary vector.

6. The method according to claim 1, characterized in that, The fuzzy logic processing described in step (3) includes: Within a preset sliding time window, the real-time reactive power deficit is compared point by point with the start-stop decision boundary vector at the corresponding time, generating a preliminary decision sequence consisting of start flags and stop flags. Using the trend statistics of continuous start-up and continuous shutdown in the preliminary sequence as fuzzy input, reasoning and defuzzification are performed through a predefined fuzzy rule base to obtain smooth decision values; The smoothing decision value is compared with preset start-up and stop-up thresholds to generate the final start-up, maintain, or stop command.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images