Power grid load intelligent prediction scheduling method and device, equipment and medium

Abstract

The application relates to a power grid load intelligent prediction scheduling method, device, equipment and medium. The method comprises the following steps: performing short-time Fourier transform on historical load data of a power grid to obtain a frequency spectrum feature vector; based on the frequency spectrum feature vector and current context information, a routing attention weight is calculated, and an adaptive learning rate multiplier and a forgetting gate threshold are generated based on concept drift significance; based on a minimum loss function, a prediction model is trained through a training sample, the adaptive learning rate multiplier and the forgetting gate threshold, so that an adaptive prediction model is obtained; historical time series data and future condition data are input into the adaptive prediction model, so that a final load point prediction and a prediction uncertainty interval are obtained; based on the final load point prediction, the prediction uncertainty interval and the routing attention weight, a scheduling optimization model is updated to generate unit combination and output planning of a target future period. The method can improve the stability and accuracy of load prediction in a variable environment.

Description

Technical Field

[0001] This invention belongs to the field of power dispatching technology, and in particular relates to a method, device, equipment and medium for intelligent prediction and dispatching of power grid load. Background Technology

[0002] With the development of power technology, renewable energy and diverse load-related applications have emerged, making the grid load exhibit strong randomness, volatility and time-varying characteristics. The load data distribution undergoes complex "conceptual drift" over time, and this drift does not evolve steadily at a constant rate, but rather exhibits non-uniform characteristics with multiple time scales and multiple driving modes.

[0003] In traditional technologies, load forecasting is mainly handled by online adaptive forecasting models. Adaptive adjustment is achieved through a homogeneous parameter update mechanism, which means that all network parameters are optimized without distinction, attempting to adapt to the dynamic changes of the load sequence with a unified parameter adjustment method.

[0004] However, when the model drastically updates its parameters to quickly adapt to short-term load changes, the neural network weights that encode long-term stable patterns are forcibly modified and overwritten, causing the model to "forget" the old patterns it has learned. This results in a persistent bias in the prediction of normal load patterns after the sudden change event ends. During the critical period of load pattern transition, external environmental signal fluctuations intensify. If the model's adaptive strategy is too sensitive, its prediction output will repeatedly jump between various historical load patterns, generating violent fluctuations without physical meaning. This will induce unnecessary frequent start-ups and shutdowns of generator units or ineffective throughput of energy storage systems, directly threatening the economy and safety of power grid operation. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and medium for intelligent prediction and scheduling of power grid load that can adaptively capture and maintain load patterns at different time scales, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for intelligent forecasting and scheduling of power grid load, including:

[0007] Historical load data, weather forecast data, and event calendar information of the power grid are acquired, and short-time Fourier transform of the historical load data of the power grid is performed based on a sliding window to obtain the spectral feature vector.

[0008] Based on spectral feature vectors and current context information, routing attention weights are calculated through a pre-trained memory router, and concept drift salience is evaluated based on spectral feature vectors to generate adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels; the current context information is obtained by encoding weather forecast data and event calendar information.

[0009] Training samples are constructed based on historical load data of the power grid, spectral feature vectors and actual load values. Based on minimizing the loss function, the prediction model based on multiple heterogeneous memory channels is trained by training samples, adaptive learning rate multiplier and forget gate threshold to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

[0010] Historical time-series data and future conditional data are input into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation to obtain the final load point prediction and prediction uncertainty interval. The historical time-series data consists of historical power grid load data, weather forecast data corresponding to the historical power grid load data, and event calendar information. The future conditional data consists of weather forecast data and event calendar information for the target future period. Multiple heterogeneous memory channels are configured to learn load patterns with different time scales and stability, and the sub-prediction results output by each memory channel are weighted and aggregated using routing attention weights.

[0011] The scheduling optimization model is updated based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight. The updated scheduling optimization model is then used to generate the unit combination and output plan for the target future time period.

[0012] In one embodiment, a short-time Fourier transform is performed on historical load data of the power grid based on a sliding window to obtain a spectral feature vector, including:

[0013] Extract load subsequences from historical load data of the power grid that meet the conditions of ending at the current time and having a preset window size, and apply Hamming windowing to the load subsequences to obtain windowed load subsequences;

[0014] Perform a Fast Fourier Transform on the windowed load subsequence and calculate the square of the spectral magnitude to obtain the time-spectrum diagram of the current sliding window;

[0015] Integrate the energy corresponding to the spectral components with frequencies less than or equal to a preset low-frequency threshold on the time-spectrum graph to obtain the low-frequency energy components;

[0016] Integrate the energy corresponding to the spectral components on the time-spectrum graph whose frequencies are greater than a preset low-frequency threshold and less than a preset high-frequency threshold to obtain the mid-frequency energy component;

[0017] Integrate the energy corresponding to the spectral components with frequencies greater than or equal to a preset high-frequency threshold on the time-spectrum graph to obtain the high-frequency energy components;

[0018] A spectral feature vector is constructed based on low-frequency energy components, mid-frequency energy components, and high-frequency energy components.

[0019] In one embodiment, the significance of concept drift is evaluated based on spectral feature vectors, and adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels are generated, including:

[0020] The concept drift significance index is calculated based on the ratio of low-frequency to high-frequency energy components in the spectral feature vector, the variation norm of historical routing attention weights, and the historical prediction mean square error of the memory channel. The formula for calculating the concept drift significance index is as follows: ,in, , and These are weighting coefficients. To prevent small positive numbers from being divided by zero, and These are the high-frequency energy components and low-frequency energy components in the spectral eigenvector, respectively. and These are the routing attention weights for the current time and the previous time, respectively. and These are the prediction mean square errors for the third and second memory channels, respectively.

[0021] For the first memory channel configured to learn long-term, slow-changing patterns, the adaptive learning rate multiplier and forgetting gate threshold are calculated based on the concept drift significance index. The formula for calculating the adaptive learning rate multiplier of the first memory channel is as follows: The formula for calculating the forgetting threshold of the first memory channel is: ,in, , , and These are the preset hyperparameters for the first memory channel. This is the significance index for concept drift;

[0022] For the second memory channel configured with a learning day cycle and baseline load pattern, the adaptive learning rate multiplier and forgetting gate threshold of the second memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the second memory channel is as follows: The formula for calculating the forgetting gate threshold of the second memory channel is: ,in and The preset hyperparameters for the second memory channel;

[0023] For the third memory channel configured to learn short-term mutations and events, the adaptive learning rate multiplier and forgetting gate threshold of the third memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the third memory channel is as follows: The formula for calculating the forgetting threshold of the third memory channel is: ,in, , , and These are the preset hyperparameters for the third memory channel.

[0024] In one embodiment, based on minimizing the loss function, a prediction model based on multiple heterogeneous memory channels is trained using training samples, an adaptive learning rate multiplier, and a forgetting gate threshold to obtain an adaptive prediction model based on multiple heterogeneous memory channels, including:

[0025] The historical load data of the power grid in the training samples is input into the prediction model based on multiple heterogeneous memory channels to obtain the initial sub-prediction results output by each memory channel;

[0026] The mean square error between the weighted aggregation result of each initial sub-prediction result and the true load value in the training sample is calculated to obtain the main prediction loss;

[0027] The real load value is decomposed into corresponding low-frequency, mid-frequency, and high-frequency true values ​​by using a preset bandpass filter.

[0028] The negative correlation coefficients between the initial sub-prediction results of the first memory channel and the low-frequency truth component, the negative correlation coefficients between the initial sub-prediction results of the second memory channel and the mid-frequency truth component, and the negative correlation coefficients between the initial sub-prediction results of the third memory channel and the high-frequency truth component are calculated respectively.

[0029] The channel isolation loss is obtained by multiplying each negative correlation coefficient by a preset isolation weight and then summing the results. The formula for calculating the channel isolation loss is as follows: ,in, The function is the Pearson negative correlation coefficient. , and These are the initial sub-prediction results for the first, second, and third memory channels, respectively. , and The low-frequency true value components, mid-frequency true value components, and high-frequency true value components are identified separately. , and For isolation weights;

[0030] The main prediction loss and channel isolation loss are weighted and summed according to a preset ratio to obtain the loss function, and the gradient of the loss function with respect to the network parameters of each memory channel is calculated through the backpropagation algorithm.

[0031] For each memory channel, the gradient is multiplied by the adaptive learning rate multiplier to obtain the adjusted gradient. The proposed update amount of the network parameters is calculated based on the adjusted gradient. It is then determined whether the ratio of the proposed update amount to the current absolute value of the network parameters exceeds the forget gate threshold to obtain the judgment result.

[0032] If the judgment result is that the proportion exceeds the forget gate threshold, the proposed update amount of the network parameters is trimmed to the range bounded by the forget gate threshold, and the network parameters of the prediction model based on multiple heterogeneous memory channels are updated based on the updated proposed update amount to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

[0033] In one embodiment, historical time-series data and future conditional data are input into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation to obtain the final load point prediction and prediction uncertainty interval, including:

[0034] Historical time-series data and future conditional data are input into an adaptive prediction model based on multiple heterogeneous memory channels to obtain the load sub-prediction sequence for each memory channel.

[0035] The load sub-prediction sequences of each memory channel are linearly weighted and aggregated based on routing attention weights to obtain the final load point prediction; the calculation formula for the final load point prediction is as follows: ,in, , and These are the load sub-prediction sequences for the first memory channel, the second memory channel, and the third memory channel, respectively. , and For routing attention weights;

[0036] Based on the historical prediction error statistics of the first, second, and third memory channels, the uncertainty variance of the load sub-prediction sequence of each memory channel is estimated, and the overall uncertainty variance of the final load point prediction is calculated according to the routing attention weight and the uncertainty variance of each memory channel.

[0037] The predicted uncertainty interval is calculated based on the overall uncertainty variance and the preset confidence level parameters.

[0038] In one embodiment, the scheduling optimization model is updated based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight. The updated scheduling optimization model is then used to generate the unit combination and output plan for the target future time period, including:

[0039] Using the final load point forecast as the baseline load curve, and combining it with a load uncertainty set constructed from the forecast uncertainty interval, a robust optimization method is employed to find the worst-case load scenario that maximizes the total system operating cost within the load uncertainty set. The optimization problem that minimizes the objective function while satisfying the constraint set under the worst-case load scenario is then solved, yielding the unit combination and output plan. The objective function is to minimize the total system operating cost, which includes generator fuel cost, unit start-up and shutdown cost, reserve capacity cost, and load deviation penalty cost. The constraint set includes system power balance constraints, upper and lower limits of each generator's output - ramp rate constraints, minimum start-up and shutdown time constraints, network power flow safety constraints, and positive and negative spinning reserve capacity constraints. The positive and negative spinning reserve capacity constraints satisfy... ,in, For positive and negative rotational reserve capacity, Based on positive and negative rotational reserve capacity This is the proportionality coefficient. This refers to the component in the routing attention weights corresponding to the third memory channel.

[0040] In one embodiment, the method further includes:

[0041] The norm of change of network parameters of the first memory channel and the second memory channel within the sliding time window is obtained. When the norm of change of the first memory channel or the second memory channel is detected to be continuously lower than the preset stability threshold for more than the preset duration, it is determined that the load mode learned by the corresponding memory channel has entered a stable state and the archiving event is triggered.

[0042] In response to an archive event, a statistical summary of the network parameters, spectral feature vector, and current context information of the memory channel currently in a stable state is encapsulated into a long-term memory pattern.

[0043] Before each training of the prediction model or before the final load point prediction, calculate the spectral feature vector and current context information, and the feature similarity with each long-term memory pattern. When the feature similarity exceeds the preset recall threshold and the prediction performance of the current prediction model on the validation set is lower than the prediction performance of the long-term memory pattern, trigger the memory retrieval-fusion event.

[0044] In response to the memory retrieval-fusion event, the parameter snapshot in the long-term memory pattern is used as the teacher model, and the corresponding memory channel in the current prediction model is used as the student model. The knowledge distillation loss is constructed and added as an additional term to the loss function.

[0045] Secondly, this application also provides a smart grid load forecasting and dispatching device, comprising:

[0046] The data module is used to acquire historical load data of the power grid, weather forecast data and event calendar information, and to perform short-time Fourier transform on the historical load data of the power grid based on a sliding window to obtain the spectral feature vector.

[0047] The boundary module is used to calculate the routing attention weights based on the spectral feature vector and the current context information through a pre-trained memory router, and to evaluate the significance of concept drift based on the spectral feature vector, generating adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels; the current context information is obtained by encoding weather forecast data and event calendar information.

[0048] The adaptive module is used to construct training samples based on historical load data of the power grid, spectral feature vectors and actual load values, and to train the prediction model based on multiple heterogeneous memory channels by minimizing the loss function, through training samples, adaptive learning rate multipliers and forget gate thresholds, to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

[0049] The prediction module is used to input historical time-series data and future conditional data into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation, so as to obtain the final load point prediction and the prediction uncertainty interval. The historical time-series data consists of historical power grid load data, weather forecast data corresponding to the historical power grid load data, and event calendar information. The future conditional data consists of weather forecast data and event calendar information for the target future period. The multiple heterogeneous memory channels are configured to learn load patterns with different time scales and stability, and the sub-prediction results output by each memory channel are weighted and aggregated using routing attention weights.

[0050] The scheduling module is used to update the scheduling optimization model based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight, and to generate the unit combination and output plan for the target future time period through the updated scheduling optimization model.

[0051] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described intelligent power grid load forecasting and scheduling methods.

[0052] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described intelligent power grid load prediction and scheduling methods.

[0053] The aforementioned intelligent power grid load forecasting and scheduling method, device, equipment, and medium acquire historical load, meteorological, and calendar information of the power grid in real time, and perform short-time Fourier transform to extract spectral feature vectors that characterize the multi-timescale fluctuations of the load sequence. Based on the spectral feature vectors and encoded context information, a pre-trained memory router dynamically calculates the routing attention weights for different learning tasks and simultaneously evaluates the significance of current concept drift to generate an adaptive learning rate multiplier and forgetting gate threshold that can differentiately regulate the plasticity and memory retention capabilities of each memory channel. Using dynamically generated regulation parameters, combined with training samples constructed from historical data and real load labels, the prediction model with multiple built-in heterogeneous memory channels is trained with the goal of minimizing the overall loss and parameter updates subject to differential constraints, thereby obtaining a predictive model that can... An adaptive non-uniform concept drift prediction model is developed. During prediction, the model integrates sub-prediction results independently calculated by each channel based on historical time series and future condition data according to routing weights, and outputs the final load point prediction and quantified uncertainty range. The dispatching system integrates load point prediction, uncertainty information, and routing weights reflecting the current pattern, updates and solves the robust optimization model, and generates a unit dispatching plan that takes into account both economy and operational safety. This achieves accurate perception and adaptive learning of the non-uniform concept drift of the power grid load, effectively solving the problems of catastrophic forgetting and model oscillation caused by memory structure imbalance in online prediction. This significantly improves the long-term stability and accuracy of load prediction in a variable environment, and provides more reliable and information-rich decision-making basis for downstream power dispatch, enhancing the overall robustness of the power grid in dealing with complex load patterns and uncertainties. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating the intelligent power grid load prediction and scheduling method of the present invention.

[0056] Figure 2 This is a structural diagram of the intelligent power grid load prediction and scheduling device of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] In one embodiment, such as Figure 1 As shown, a method for intelligent forecasting and scheduling of power grid load is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0059] S101. Obtain historical load data, weather forecast data, and event calendar information of the power grid, and perform short-time Fourier transform on the historical load data of the power grid based on a sliding window to obtain the spectral feature vector.

[0060] Indicatively, historical load data of the power grid covers core load characterization data such as active power load and reactive power load values ​​of each node and region of the power grid within a preset time span; weather forecast data includes historical observation and forecast values ​​of meteorological elements such as temperature, humidity, wind speed, precipitation, and light intensity in the target area; event calendar information covers non-meteorological events that may affect the power grid load, such as statutory holidays, major social events, power grid maintenance plans, and industrial production schedules.

[0061] Furthermore, the sliding window length covers at least one complete load cycle feature, and the sliding step size is consistent with the time granularity of data acquisition. A short-time Fourier transform is performed on the historical load data of the power grid within the sliding window to extract the spectral feature vector, i.e. ,in The load data sequence within the sliding window. For time indexing; This is the time offset from the center of the window; For frequency variables; For window functions, you can choose Hanning window, Hamming window or rectangular window, etc., which are suitable for non-stationary time series analysis; The imaginary unit satisfies The above transformation yields amplitude and phase information in frequency dimension. This amplitude and phase information is then regularized according to a preset frequency range to form a fixed-dimensional spectral feature vector. Each dimension of this vector corresponds to the load spectral characteristics within a specific frequency range, which can effectively characterize the periodicity and fluctuation characteristics of load data in different frequency dimensions.

[0062] S102. Based on the spectral feature vector and the current context information, the routing attention weights are calculated through a pre-trained memory router, and the significance of concept drift is evaluated based on the spectral feature vector. Adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels are generated respectively. The current context information is obtained by encoding weather forecast data and event calendar information.

[0063] Specifically, weather forecast data and event calendar information are encoded to obtain current context information. For continuous meteorological elements in the weather forecast data, normalization is used to map them to numerical ranges, while for discrete meteorological elements, one-heat encoding is used to convert them into binary feature vectors. For event calendar information, multi-dimensional encoding rules are constructed based on event type and impact level, converting the event's time attribute, regional attribute, and impact type attribute into numerical features. Finally, the normalized weather forecast feature vector and the encoded event calendar feature vector are concatenated to obtain a current context information feature vector with unified dimensions.

[0064] Furthermore, the spectral feature vector and the current context information feature vector are input into a pre-trained memory router to fuse the two types of feature vectors. The fusion method uses element-wise multiplication combined with fully connected layer mapping to obtain the fused feature vector. The routing attention weights are calculated using the attention calculation module of the memory router. ,in, The weight matrix for the attention weight calculation layer. For bias vectors, The function is used to normalize the calculation results to ensure that the sum of each weight value is 1. The dimension of the routing attention weight is consistent with the number of heterogeneous memory channels, and each weight value corresponds to the attention of a heterogeneous memory channel.

[0065] Optionally, concept drift significance can be evaluated based on spectral feature vectors. This is achieved by calculating the cosine similarity between spectral feature vectors corresponding to different sliding windows, using the reciprocal of this similarity as the base value for concept drift significance, and then correcting it by incorporating the time-series rate of change of the load data. The resulting concept drift significance is then calculated. ,in, For the first The spectral feature vectors corresponding to each sliding window For the first The spectral feature vectors corresponding to each sliding window This is the cosine similarity calculation function. These are weighting coefficients used to balance the contributions of spectral similarity and temporal rate of change. The time-series rate of change of load data is calculated as the ratio of the difference in load values ​​at adjacent time points to the mean. Further, based on the significance of concept drift... Generate multiple heterogeneous memory channels, each with its own adaptive learning rate multiplier and forget gate threshold. The adaptive learning rate multiplier follows... ,in, The drift sensitivity coefficient, For the first The base learning rate multiplier for each memory channel. Indexing heterogeneous memory channels; forget gate threshold follows ,in, This is the threshold adjustment coefficient. For the first The basic forgetting gate threshold for each memory channel.

[0066] S103. Based on historical load data of the power grid, spectral feature vectors and actual load values, training samples are constructed. Based on minimizing the loss function, the prediction model based on multiple heterogeneous memory channels is trained by training samples, adaptive learning rate multiplier and forget gate threshold to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

[0067] For example, the load data processed by the sliding window, the corresponding spectral feature vector, and the actual load value at the same time dimension are mapped one-to-one. Each training sample includes an input feature part and a label part. The input feature part is composed of the time-series features and spectral feature vector of the historical load data of the power grid, and the label part is the actual load value at the corresponding time point. Optionally, the loss function that aims to minimize the error between the predicted value and the actual load value can be the mean squared error loss function, i.e. ,in, The number of training samples, For the first The actual load value corresponding to each training sample. For the prediction model to the first The predicted load value output by each training sample.

[0068] Specifically, training samples are input into a prediction model based on multiple heterogeneous memory channels. The model's input layer performs feature mapping on the training samples, and the mapped features are then input into the multiple heterogeneous memory channels. During parameter updates, each heterogeneous memory channel multiplies its corresponding adaptive learning rate multiplier with the base learning rate to obtain the actual learning rate for that channel. ,in, For the first The actual learning rate of each heterogeneous memory channel This sets the base learning rate for the model. Optionally, a forget gate structure is set for each heterogeneous memory channel. The output of the forget gate is calculated from the forget gate threshold and the hidden states within the channel. ,in, It is the Sigmoid activation function. For the first The hidden state of a heterogeneous memory channel Here is the weight matrix for the forget gate. For the bias term of the forget gate, This is the output value of the forget gate, which controls the proportion of historical hidden states retained within the channel. The closer the value is to 1, the more hidden historical states are retained and the lower the degree of forgetting; conversely, the lower the value is, the higher the degree of forgetting. Using the backpropagation algorithm, based on the objective of minimizing the loss function, the model parameters are iteratively updated by combining the actual learning rate of each heterogeneous memory channel with the forgetting gate threshold. This iterative process continues until the loss function value converges to a preset threshold or reaches a preset number of iterations, ultimately yielding an adaptive prediction model based on multiple heterogeneous memory channels.

[0069] S104. Input historical time-series data and future conditional data into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation to obtain the final load point prediction and prediction uncertainty interval. The historical time-series data consists of historical power grid load data, weather forecast data corresponding to the historical power grid load data, and event calendar information. The future conditional data consists of weather forecast data and event calendar information for the target future period. Multiple heterogeneous memory channels are configured to learn load patterns with different time scales and stability, and the sub-prediction results output by each memory channel are weighted and aggregated using routing attention weights.

[0070] Furthermore, the historical load data of the power grid, the weather forecast data and event calendar information corresponding to the historical load data of the power grid are time-aligned to ensure that the data under the same timestamp correspond one-to-one. The aligned multi-source data is converted into a numerical feature vector of a unified dimension to obtain historical time series data. The future condition data uses the target future period as the time base, obtains the meteorological element forecast values ​​and event information corresponding to the period, and converts them into numerical feature vectors using the same encoding method as the historical time series data.

[0071] Historical time-series data and future conditional data are input into an adaptive prediction model based on multiple heterogeneous memory channels, performing isolated forward propagation. Isolated forward propagation means that each heterogeneous memory channel maintains independent parameter updates during feature processing; information exchange between different memory channels occurs only at the output layer, with no parameter sharing or interference. Multiple heterogeneous memory channels are configured to learn load patterns with different time scales and stability. For example, some channels are adapted to short-term load patterns with a 24-hour cycle, with a high parameter update frequency and sensitivity to short-term fluctuations; others are adapted to medium- to long-term stable load patterns with monthly or quarterly scales, with a lower parameter update frequency and a higher forgetting threshold to reduce forgetting of long-term patterns. Each heterogeneous memory channel outputs a corresponding sub-prediction result based on the input data. Furthermore, the routing attention weights are used to weight and aggregate the results of each sub-prediction, i.e. ,in, This represents the total number of heterogeneous memory channels. For the first The routing attention weights corresponding to each heterogeneous memory channel. This represents the final load point prediction. The calculation of the prediction uncertainty interval is based on the variance of each sub-prediction result and the routing attention weights. The weighted variance is obtained as The upper and lower limits of the prediction uncertainty interval are determined based on the weighted variance, with the upper limit being... The lower limit of the interval is ,in The quantiles are the confidence levels.

[0072] S105. Update the scheduling optimization model based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight, and generate the unit combination and output plan for the target future period through the updated scheduling optimization model.

[0073] In a schematic representation, the final load point forecast is used as the basic load demand input for the dispatch optimization model, the upper and lower limits of the forecast uncertainty interval are used as the boundary constraints for load demand fluctuations, and the routing attention weight is used as a representation of the importance of load patterns at different time scales, all integrated into the objective function and constraints of the dispatch optimization model. The objective function of the dispatch optimization model focuses on the economy and security of power grid operation. ,in, The operating costs of the generator set include fuel costs, start-up and shutdown costs, etc. The dispatch cost of the energy storage system; The risk cost is calculated based on the predicted uncertainty interval, and this risk cost is positively correlated with the width of the predicted uncertainty interval; the wider the interval, the higher the risk cost. The constraints of the scheduling optimization model include upper and lower limits of unit output, unit ramp rate constraints, grid power balance constraints, and energy storage system charging and discharging constraints. During the constraint construction process, the constraint relaxation corresponding to different time scale load patterns is adjusted by incorporating routing attention weights.

[0074] The updated scheduling optimization model is solved using optimization algorithms such as linear programming and mixed integer programming. The solution process satisfies all constraints and minimizes the objective function value, ultimately generating the unit combination and output plan for the target future time period. The unit combination scheme clarifies the start-up and shutdown status of each generator unit in the target future time period, and the output plan clarifies the active power and reactive power output values ​​of each operating unit at different time points.

[0075] In the aforementioned intelligent load forecasting and scheduling method for power grids, a sliding window is used to perform a short-time Fourier transform on historical load data to extract spectral feature vectors. Routing attention weights are calculated based on a pre-trained memory router. Simultaneously, the significance of concept drift is evaluated to generate adaptive learning rate multipliers and forget gate thresholds for multiple heterogeneous memory channels. With the goal of minimizing the loss function, the adaptive learning rate multipliers and forget gate thresholds are combined to train a multi-heterogeneous memory channel prediction model for learning load patterns with different time scales and stability. This yields an adaptive multi-heterogeneous memory channel prediction model. Isolated forward propagation and weighted aggregation of sub-prediction results from each memory channel using routing attention weights are implemented to obtain the final load point prediction and prediction uncertainty interval. This updates the scheduling optimization model and generates unit combinations and output plans for the target future time period. This effectively solves the load forecasting problem caused by non-uniform concept drift, achieves accurate adaptive load forecasting, improves the smoothness and physical rationality of the prediction results, provides more valuable input information for power grid scheduling, and ultimately ensures the economic efficiency and safety of power grid operation.

[0076] In one embodiment, a short-time Fourier transform is performed on historical load data of the power grid based on a sliding window to obtain a spectral feature vector, including:

[0077] S11. Extract a load subsequence from the historical load data of the power grid that meets the condition of ending at the current time and having a preset window size. Apply a Hamming window to the load subsequence to obtain a windowed load subsequence.

[0078] In a schematic representation, taking the current forecast time as the end point of the time axis, continuous data points are extracted in reverse order according to a preset time resolution to form a load subsequence of a preset window size. The extracted load subsequence is then windowed using a Hamming window to obtain the windowed load subsequence. ,in, This is the index of the sampling points within the window, with a value range of [value range missing]. ; The length of the Hanming window should be consistent with the preset window size.

[0079] S12. Perform a fast Fourier transform on the windowed load subsequence and calculate the square of the spectral magnitude to obtain the time spectrum of the current sliding window.

[0080] Performing a Discrete Fast Fourier Transform on the windowed load subsequence converts the load signal from the time domain to the frequency domain. ,in, These are complex spectral values ​​in the frequency domain. This is the frequency point index, with a value range of [value range missing]. ; For the windowed load subsequence Data from each sampling point; For the first Hanming window Window function value; The imaginary unit satisfies ; The length of the windowed load subsequence is the window length. Furthermore, the square of the spectral magnitude is... ,in, Complex spectral values The real part, Complex spectral values The imaginary part. The time-frequency spectrum is indexed by frequency points. With time frames as the vertical axis and time frames as the horizontal axis, The amplitude is used to visually represent the energy distribution characteristics of the load signal at different frequencies and time dimensions.

[0081] S13. Integrate the energy corresponding to the spectral components with frequencies less than or equal to the preset low-frequency threshold on the time-frequency spectrum to obtain the low-frequency energy components.

[0082] Preset low frequency threshold The corresponding load signal has a period greater than the upper limit of a preset long-term time scale. For example, the frequency on the time spectrum satisfies... Integrating the energy corresponding to the spectral components, i.e. ,in, It is a low-frequency energy component. For continuous frequency variables, The low-frequency energy component obtained is the spectral power density function, used to characterize the strength of the long-term stable trend in the load sequence.

[0083] S14. Integrate the energy corresponding to the spectral components with frequencies greater than the preset low-frequency threshold and less than the preset high-frequency threshold on the time-frequency spectrum to obtain the mid-frequency energy component.

[0084] Preset high frequency threshold ,satisfy The corresponding load signal has a frequency lower limit whose period is less than a preset short-term time scale. For example, the frequency on the time spectrum satisfies... Integrating the energy corresponding to the spectral components, i.e. ,in, The mid-frequency energy component is used to obtain the mid-frequency energy component that characterizes the intensity of the daily cycle pattern and related variants in the load sequence.

[0085] S15. Integrate the energy corresponding to the spectral components with frequencies greater than or equal to the preset high-frequency threshold on the time-frequency spectrum to obtain the high-frequency energy components.

[0086] The frequencies on the time spectrum satisfy Integrating the energy corresponding to the spectral components, i.e. ,in, It is a high-frequency energy component. The highest analysis frequency, whose value is determined by the sampling frequency of the load data, satisfies the Nyquist sampling theorem. , The sampling frequency is used to obtain high-frequency energy components to characterize the intensity of short-term fluctuations, random noise, and sudden load changes caused by sudden events in the load sequence.

[0087] S16. Construct a spectral feature vector based on low-frequency energy components, mid-frequency energy components, and high-frequency energy components.

[0088] Based on the obtained low-frequency energy components Mid-frequency energy component and high-frequency energy components Constructing spectral feature vectors .

[0089] In one embodiment, the significance of concept drift is evaluated based on spectral feature vectors, and adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels are generated, including:

[0090] S21. Based on the ratio of low-frequency energy components to high-frequency energy components in the spectral feature vector, the variation norm of historical routing attention weights, and the historical prediction mean square error of the memory channel, the concept drift significance index is calculated; the formula for calculating the concept drift significance index is as follows: ,in, , and These are weighting coefficients. To prevent small positive numbers from being divided by zero, and These are the high-frequency energy components and low-frequency energy components in the spectral eigenvector, respectively. and These are the routing attention weights for the current time and the previous time, respectively. and These are the prediction mean square errors for the third and second memory channels, respectively.

[0091] Indicative, satisfying And all of them are non-negative real numbers; This is used to quantify the magnitude of changes in routing decisions; a larger norm indicates a more drastic shift in load patterns. The mean square error of the prediction for the third memory channel, which focuses on learning short-term mutations and event load patterns, reflects the ability to adapt to sudden load changes. The second memory channel represents the prediction mean square error, which focuses on learning the daily cycle and baseline load patterns. Its error reflects the fitting ability to stable load patterns. The ratio of the two is used to determine whether the error originates from insufficient adaptation to burst patterns or fitting bias in stable patterns. Finally, a significance index that comprehensively reflects the degree of conceptual drift in the load sequence is obtained through weighted summation. .

[0092] S22. For the first memory channel configured to learn long-term, slow-changing patterns, calculate the adaptive learning rate multiplier and forgetting gate threshold corresponding to the first memory channel based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the first memory channel is as follows: The formula for calculating the forgetting threshold of the first memory channel is: ,in, , , and These are the preset hyperparameters for the first memory channel. This is the significance index for concept drift.

[0093] For example, The drift sensitivity coefficient of the first memory channel is used to control the response strength of the learning rate to concept drift. Its value is preset based on the stability requirements of the long-term load mode. This is the minimum learning rate multiplier for the first memory channel, used as a lower threshold to ensure necessary updates to channel parameters, preventing the learning rate from becoming too low due to drastic concept drift and failing to adapt to the slow evolution of long-term trends. Through Function constraints ensure that the learning rate multiplier is not less than This suppresses the parameter update rate of long-term memory channels, thus preventing long-term stable knowledge from being disturbed by short-term mutations. The basic forgetting threshold for the first memory channel is a benchmark value to ensure normal memory renewal; This is the threshold adjustment coefficient for the first memory channel, used to control the sensitivity of the forgetting gate threshold to concept drift. The more significant the concept drift, the lower the sensitivity. The larger the threshold, the higher the forgetting threshold. The larger the value, the lower the tolerance of the memory channel to parameter rewriting. It needs to meet a higher range of change conditions before it is allowed to be updated, thereby strengthening the protection of long-term stable memory and avoiding "catastrophic forgetting".

[0094] S23. For the second memory channel configured with a learning day cycle and baseline load mode, the adaptive learning rate multiplier and forget gate threshold of the second memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the second memory channel is as follows: The formula for calculating the forgetting gate threshold of the second memory channel is: ,in and These are the preset hyperparameters for the second memory channel.

[0095] For example, This is the drift sensitivity coefficient for the second memory channel, with a value between that of the first and third memory channels, used to balance the response to and suppression of drift. The daily cycle pattern's response to drift needs to be lower than that of the short-term abrupt change channel but higher than that of the long-term trend channel. The square root operation mitigates the impact of drift on the learning rate, ensuring that the parameter update rate can adapt to the small dynamic changes in the daily cycle pattern without over-updating due to severe drift. Since the daily cycle pattern is the core stable component of the grid load, its variation is relatively gradual, and there is no need to adjust the forget gate threshold according to the concept drift dynamics. A fixed threshold can ensure the persistence and stability of the daily cycle memory, while adapting to subtle variations in the daily cycle pattern through appropriate parameter updates.

[0096] S24. For the third memory channel configured to learn short-term mutations and events, the adaptive learning rate multiplier and forgetting gate threshold of the third memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the third memory channel is as follows: The formula for calculating the forgetting threshold of the third memory channel is: ,in, , , and These are the preset hyperparameters for the third memory channel.

[0097] For example, This is the drift sensitivity coefficient for the third memory channel, which is higher than that for the first and second memory channels to ensure a strong response to concept drift. This is the maximum learning rate multiplier for the third memory channel, used to limit the upper limit of the learning rate and avoid parameter oscillations or overfitting due to an excessively high learning rate. The more significant the concept drift, the better. The larger the threshold, the higher the forgetting threshold. The smaller the value, the lower the tolerance threshold for parameter rewriting, allowing the third memory channel to quickly forget short-term historical memories unrelated to the current mutation mode, freeing up memory space to efficiently learn new mutation features, and improving the speed of adaptation to sudden load changes.

[0098] In one embodiment, based on minimizing the loss function, a prediction model based on multiple heterogeneous memory channels is trained using training samples, an adaptive learning rate multiplier, and a forgetting gate threshold to obtain an adaptive prediction model based on multiple heterogeneous memory channels, including:

[0099] S31. Input the historical load data of the power grid in the training sample into the prediction model based on multiple heterogeneous memory channels to obtain the initial sub-prediction results output by each memory channel.

[0100] Indicatively, each heterogeneous memory channel, based on its own pre-configured network structure and initial parameters, independently extracts features and learns patterns from the input historical load data of the power grid. Each memory channel, according to its assigned load pattern learning task, maps the input data to the corresponding prediction output through forward propagation operations in its internal network layer. This results in the initial sub-prediction results output by the first, second, and third memory channels, respectively. Each initial sub-prediction result maintains consistency with the time dimension of the training samples, corresponding to the predicted values ​​of three load patterns: long-term trend, daily cycle baseline, and short-term abrupt change.

[0101] S32. Calculate the mean square error between the weighted aggregation result of each initial sub-prediction result and the true load value in the training sample to obtain the main prediction loss.

[0102] For example, the master prediction loss is ,in Forecast the loss. The time step length of a single sample in the training samples. For the first The weighted aggregated prediction values ​​at each time step. For the first The actual load value at each time step.

[0103] S33. The real load value is decomposed into the corresponding low-frequency true value component, mid-frequency true value component and high-frequency true value component by a preset bandpass filter.

[0104] Similarly, the passband frequency range of the low-frequency bandpass filter is... It is used to separate the low-frequency true value component corresponding to the long-term stable trend in the actual load value; the passband frequency range of the intermediate frequency bandpass filter is... This is used to separate the intermediate frequency true components corresponding to the daily cycle and baseline load pattern; the passband frequency range of the high-frequency bandpass filter is... This is used to separate the high-frequency truth components corresponding to short-term mutations and events, and to decompose the true load value into mutually independent low-frequency truth components. Intermediate frequency true value components and high-frequency true components .

[0105] S34. Calculate the negative correlation coefficient between the initial sub-prediction result of the first memory channel and the low-frequency truth component, the negative correlation coefficient between the initial sub-prediction result of the second memory channel and the mid-frequency truth component, and the negative correlation coefficient between the initial sub-prediction result of the third memory channel and the high-frequency truth component, respectively.

[0106] For example, the Pearson correlation coefficient between the initial sub-prediction result of each memory channel and the corresponding true value component of the frequency band is calculated, i.e., the initial sub-prediction result of the first memory channel. With low-frequency true components The Pearson correlation coefficient, the initial sub-prediction results of the second memory channel. With intermediate frequency true value components The Pearson correlation coefficient, the initial sub-prediction results of the third memory channel. With high-frequency true value components The Pearson correlation coefficient was calculated, and the negative values ​​of each correlation coefficient were taken to obtain three corresponding negative correlation coefficients.

[0107] S35. Multiply each negative correlation coefficient by a preset isolation weight and sum them to obtain the channel isolation loss; the formula for calculating the channel isolation loss is as follows: ,in, The function is the Pearson negative correlation coefficient. , and These are the initial sub-prediction results for the first, second, and third memory channels, respectively. , and The low-frequency true value components, mid-frequency true value components, and high-frequency true value components are identified separately. , and For isolation weights.

[0108] Specifically, The function for calculating the Pearson correlation coefficient takes two numerical sequences of equal length as input and outputs their linear correlation coefficient. Forecasting corresponding to long-term trend load patterns; Forecasting corresponding to daily cycles and baseline load patterns; Prediction of short-term mutations and event load patterns; , and The isolation weights are used to adjust the strength of the isolation constraints for each channel, encouraging each memory channel to focus on learning the load patterns of its own frequency band. When the sub-prediction of a channel is more strongly correlated with the corresponding true value component, the contribution of the corresponding term of that channel to the loss is smaller, and vice versa, thus penalizing "cross-band" learning behavior.

[0109] S36. The master prediction loss and channel isolation loss are weighted and summed according to a preset ratio to obtain the loss function, and the gradient of the loss function with respect to the network parameters of each memory channel is calculated through the backpropagation algorithm.

[0110] Furthermore, the total loss function is ,in For the total loss function, Forecast the loss. For the loss of channel isolation, A preset scaling factor is used to balance the accuracy of the main prediction with the channel isolation effect. Its value is determined based on the convergence of model training and the optimization objective of prediction performance. The backpropagation algorithm is used to calculate the gradient of the total loss function with respect to the network parameters of each memory channel. The gradient calculation process follows the chain rule, starting from the total loss function and deriving the partial derivatives of the parameters of each network layer layer by layer to obtain an independent gradient vector for each memory channel. This gradient vector reflects the direction and degree of influence of the changes in each parameter on the total loss.

[0111] S37. For each memory channel, multiply the gradient by the adaptive learning rate multiplier to obtain the adjusted gradient, calculate the proposed update amount of the network parameters based on the adjusted gradient, and determine whether the ratio of the proposed update amount to the current absolute value of the network parameters exceeds the forget gate threshold to obtain the judgment result.

[0112] For each memory channel, its corresponding gradient vector is element-wise multiplied by the adaptive learning rate multiplier for that channel to obtain an adjusted gradient vector. This adjusted gradient vector controls the step size of parameter updates, ensuring that the parameter update rate for each channel adapts to the current concept drift state. Based on the adjusted gradient vector, the proposed update amount of the network parameters is calculated. The proposed update amount is calculated as a vector operation along the gradient descent direction; that is, the proposed update amount equals the product of the adjusted gradient vector and the preset gradient descent direction coefficient. The ratio of the absolute value of each element in the proposed update amount to the absolute value of the corresponding network parameter's current value is calculated. This ratio quantifies the relative magnitude of the proposed parameter update. This ratio is then compared with the forget gate threshold for that memory channel to determine whether the relative magnitude of the proposed update exceeds the forget gate threshold, resulting in a binary judgment result.

[0113] S38. If the judgment result is that the proportion exceeds the forget gate threshold, the proposed update amount of the network parameters is trimmed to the range bounded by the forget gate threshold, and the network parameters of the prediction model based on multiple heterogeneous memory channels are updated based on the updated proposed update amount to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

[0114] If the result indicates that the ratio exceeds the forget gate threshold, it means that the current parameter update amount may cause excessive rewriting of the core memory of this channel, and the update amount needs to be pruned. Specifically, the pruning operation is performed on each element in the update amount. The range of values ​​after clipping is ,in This is the forgetting threshold for this memory channel. The corresponding current network parameter value, i.e. ,in, This represents the pruned element to be updated. Based on the pruned element to be updated, the network parameters of this memory channel are updated, with the gradient descent direction set to positive, and the update formula is: After all memory channels have completed parameter updates, an adaptive prediction model based on multiple heterogeneous memory channels is obtained. This model achieves accurate learning and memory protection for load patterns at different time scales through differentiated parameter updates and constraints.

[0115] In one embodiment, historical time-series data and future conditional data are input into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation to obtain the final load point prediction and prediction uncertainty interval, including:

[0116] S41. Input historical time series data and future conditional data into an adaptive prediction model based on multiple heterogeneous memory channels to obtain the load sub-prediction sequence of each memory channel.

[0117] In a schematic manner, historical time-series data and future conditional data are uniformly organized according to preset feature dimensions and time granularity to ensure that the two types of data are precisely aligned on the time axis, and that the feature dimensions completely match the input layer dimensions of the adaptive prediction model based on multiple heterogeneous memory channels. Each heterogeneous memory channel performs independent, isolated forward propagation processing on the input data based on its own updated network parameters and internal state. During the isolated forward propagation process, each channel only uses its own network structure and parameters to perform feature extraction, pattern matching, and prediction operations on the data, with no parameter sharing or feature interaction between channels. Specifically, the first memory channel focuses on the long-term trend features contained in the historical time-series data and the long-term influencing factors in the future conditional data, outputting a load sub-prediction sequence reflecting the long-term stable load pattern; the second memory channel focuses on extracting daily cycle patterns and baseline load features, outputting a load sub-prediction sequence adapted to the daily cycle fluctuation pattern; the third memory channel quickly captures short-term mutation signals and event-driven features in the data, outputting a load sub-prediction sequence responding to short-term dynamic changes, ultimately obtaining the load sub-prediction sequences corresponding to the three memory channels, with the time dimension of each sequence consistent with the target future period.

[0118] S42. Based on the routing attention weights, the load sub-prediction sequences of each memory channel are linearly weighted and aggregated to obtain the final load point prediction; the calculation formula for the final load point prediction is as follows: ,in, , and These are the load sub-prediction sequences for the first memory channel, the second memory channel, and the third memory channel, respectively. , and This refers to the routing attention weights.

[0119] For example, satisfying Each component is a non-negative real number, representing the contribution weight of the load sub-prediction sequences of the first memory channel, the second memory channel, and the third memory channel under the current load mode.

[0120] S43. Based on the historical prediction error statistics of the first memory channel, the second memory channel and the third memory channel, estimate the uncertainty variance of the load sub-prediction sequence of each memory channel, and calculate the overall uncertainty variance of the final load point prediction according to the routing attention weight and the uncertainty variance of each memory channel.

[0121] Specifically, for each memory channel, the most recent preset number of historical prediction periods are selected as a statistical window, and the error value between the load sub-prediction sequence output by the channel within this window and the actual load sequence of the same period is extracted. ,in For channel index, For historical time step indexing, the prediction error variance of each channel is calculated using the unbiased variance estimation formula, i.e., the uncertainty variance of each memory channel. For example... ,in, For the first Uncertainty variance of each memory channel To count the number of historical time steps within the window, This represents the mean of the prediction error for this channel within the statistical window. The overall uncertainty variance of the final load point prediction is... ,in, The overall uncertainty variance for the final load point prediction is calculated as follows: since the prediction errors of each memory channel are independent of each other, only the weighted sum of squares of the variances of each channel is retained, with the weights being the squares of the corresponding routing attention weights. This ensures that the overall uncertainty variance can accurately reflect the combined impact of the prediction uncertainty of each channel on the final prediction result. The uncertainty variance of the first memory channel. The variance of the uncertainty in the second memory channel. This represents the uncertainty variance of the third memory channel.

[0122] S44. Calculate the predicted uncertainty interval based on the overall uncertainty variance and the preset confidence level parameters.

[0123] Optionally, the preset confidence level parameter corresponds to the quantile of the standard normal distribution. quantiles The value of is determined by a preset confidence level, which corresponds one-to-one with a quantile, and is used to characterize the probability that the actual load value falls within the prediction uncertainty interval. The prediction uncertainty interval is... ,in, To predict the uncertainty interval, For the final load point prediction sequence, To set the standard normal distribution quantiles corresponding to the preset confidence level, This represents the overall forecast standard deviation.

[0124] In one embodiment, the scheduling optimization model is updated based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight. The updated scheduling optimization model is then used to generate the unit combination and output plan for the target future time period, including:

[0125] S51. Using the final load point forecast as the baseline load curve, and combining it with the load uncertainty set constructed from the forecast uncertainty interval, a robust optimization method is used to find the worst-case load scenario within the load uncertainty set that maximizes the total system operating cost. The optimization problem that satisfies the constraint set and minimizes the objective function under the worst-case load scenario is then solved, yielding the unit combination and output plan. The objective function is to minimize the total system operating cost, which includes generator fuel cost, unit start-up and shutdown cost, reserve capacity cost, and load deviation penalty cost. The constraint set includes system power balance constraints, upper and lower limits of each generator's output - ramp rate constraints, minimum start-up and shutdown time constraints, network power flow safety constraints, and positive and negative spinning reserve capacity constraints. The positive and negative spinning reserve capacity constraints satisfy... ,in, For positive and negative rotational reserve capacity, Based on positive and negative rotational reserve capacity This is the proportionality coefficient. This refers to the component in the routing attention weights corresponding to the third memory channel.

[0126] Schematic representation: For each time step of the target future period, the possible range of actual load values ​​is defined by the prediction uncertainty interval corresponding to that time step. The load uncertainty set is the set of all load curves that satisfy the value constraints at each time step. ,in, For a certain feasible load curve in the th Load value at time step For the first The final load point prediction at the time step. For the first The overall forecast standard deviation of the time step, The total number of time steps for the target future period. The quantiles corresponding to the preset confidence level.

[0127] Furthermore, robust optimization methods are employed on load uncertainty sets. The system seeks the worst-case load scenario that maximizes the total operating cost of the system. This scenario is characterized by the maximum impact of fluctuating load demand combinations on the grid operating cost. By solving the optimization problem under this scenario, the scheduling scheme can be ensured to be feasible and economical under all possible load scenarios.

[0128] The objective function of the optimization problem is to minimize the total operating cost of the system. For example, ,in, The total operating cost of the system; The cost of power generation fuel refers to the expenses incurred by each generator unit in consuming fuel during the dispatch cycle, which is directly related to the unit's output level, fuel type, and consumption rate. The start-up and shutdown cost of the generator unit refers to the fixed costs incurred when the generator unit starts from a stopped state or stops from an operating state, including equipment wear and tear, start-up fuel consumption and other related expenses. The cost of standby capacity refers to the cost incurred for the standby capacity reserved to cope with load fluctuations and equipment failures, covering the cost of occupying various types of standby such as spinning standby and cold standby; Load deviation penalty cost refers to the penalty cost incurred when there is a deviation between the actual load and the predicted load due to untimely scheduling adjustments or insufficient resource allocation. The larger the deviation, the higher the penalty cost.

[0129] For example, system power balance constraints require that at each time step, the sum of the total output of all generator units, the charging and discharging power of the energy storage system, and the adjustment amount of interruptible loads must be balanced with the actual load value of the current time step to ensure grid frequency stability; upper and lower limit constraints on the output of each generator unit define the minimum and maximum output range allowed for each unit during operation, which is determined by the physical performance and operating procedures of the unit; ramp rate constraints limit the maximum variation in the output of the generator unit per unit time to prevent sudden output changes from impacting the unit equipment and grid stability; minimum start-up and shutdown time constraints specify the minimum time that the generator unit needs to run continuously after startup and the minimum time that it needs to remain in a shutdown state after shutdown, avoiding increased equipment wear caused by frequent start-up and shutdown of units; network power flow safety constraints ensure that the power transmission of each branch of the grid does not exceed its safe current carrying capacity to prevent safety accidents caused by line overload; positive and negative spinning reserve capacity constraints are used to ensure the grid's ability to cope with load fluctuations, i.e. ,in, The total positive and negative rotating reserve capacity required by the system refers to the sum of the output that can be quickly increased and decreased from all operating units; The basic positive and negative rotating reserve capacity is the minimum reserve capacity standard preset based on the normal operating needs of the power grid; This is a scaling factor used to adjust the influence of routing attention weights on reserve capacity; its value is calibrated through training in historical scheduling scenarios. This is the component corresponding to the third memory channel in the routing attention weights. The magnitude of this component directly reflects the proportion of short-term mutations and event-driven components in the current load pattern. A larger value indicates that the load faces a higher risk of sudden changes, and the grid's ability to withstand fluctuations needs to be improved by increasing the spinning reserve capacity.

[0130] The constrained minimization problem described above is solved using a numerical optimization algorithm. The solution process traverses the worst-case scenario within the load uncertainty set to ensure that the optimization result satisfies all constraints and minimizes the objective function value under that scenario. The final solution yields the unit combination and output plan for the target future time period. The unit combination specifies the start-up and shutdown status of each generator unit at each time step, while the output plan determines the specific active power output setpoints of each operating unit at each time step. It also includes the charging and discharging scheduling plan for the energy storage system and the dispatching arrangements for interruptible loads, providing a concrete execution basis for the safe and economical operation of the power grid.

[0131] In one embodiment, the method further includes:

[0132] S61. Obtain the norm of change of the network parameters of the first memory channel and the second memory channel within the sliding time window. When the norm of change of the first memory channel or the second memory channel is continuously lower than the preset stability threshold for more than the preset duration, determine that the load mode learned by the corresponding memory channel has entered a stable state and trigger the archiving event.

[0133] Indicatively, the network parameters include the weight matrices and bias vectors of each layer, and the parameter variation norm is calculated using the L2 norm, i.e. ,in For the first The norm of parameter variation for each memory channel within a sliding time window; For the first The total number of network parameters for each memory channel, covering all elements of the weight matrix and the bias vector; For the first The memory channel in the first Network parameter values ​​during the training cycle; For the first The memory channel in the first Network parameter values ​​during the training period.

[0134] The system monitors the parameter change norms of the first and second memory channels in real time. When the change norm of either channel is continuously lower than the preset stability threshold and the duration reaches the preset duration, it is determined that the load pattern learned by that channel has entered a stable state, that is, the pattern features no longer change significantly with training iterations, and the system automatically triggers the archiving event.

[0135] S62. In response to the archive event, encapsulate the network parameter snapshot, spectral feature vector, and statistical summary of the current context information of the memory channel that is currently in a stable state into a long-term memory pattern.

[0136] Specifically, the parameter snapshot is a complete copy of all network parameters in the current stable state of the memory channel, including the weight matrices, bias vectors, and internal state variables of each hidden layer, ensuring that the long-term memory mode can completely reproduce the model's predictive ability in the stable state. The statistical summary of the spectral feature vectors is a statistical characterization of all spectral feature vectors within the sliding time window. The mean and variance of each dimension are calculated to form a six-dimensional statistical feature vector, used to characterize the load spectrum distribution characteristics corresponding to the stable mode. The statistical summary of the current context information covers weather forecast data and event calendar information within the sliding time window. The statistical summary of meteorological data includes the mean and extreme values ​​of continuous meteorological elements such as average temperature and humidity, while the statistical summary of event calendar information includes event type distribution and event frequency, forming multi-dimensional contextual statistical features.

[0137] S63. Before each training of the prediction model or before the prediction of the final load point, calculate the spectral feature vector and the current context information, and the feature similarity with each long-term memory pattern. When the feature similarity exceeds the preset recall threshold and the prediction performance of the current prediction model on the validation set is lower than the prediction performance of the long-term memory pattern, trigger the memory retrieval-fusion event.

[0138] Furthermore, before each training iteration of the prediction model or before the final load point prediction, a fused feature vector of the current spectral feature vector and the current context information is calculated. The fusion method involves concatenating the spectral feature vector and the encoded context information feature vector along their dimensions to obtain the current feature vector. Simultaneously, statistical summaries of spectral features and contextual information stored in each long-term memory pattern are extracted, and feature vectors for each long-term memory pattern are constructed according to the same fusion rules. .

[0139] Calculate the feature similarity between the current feature vector and the feature vectors of each long-term memory pattern, using cosine similarity as the similarity metric. ,in, This is the cosine similarity calculation function. A preset recall threshold is used to determine the degree of feature matching; when the feature similarity exceeds this threshold, it indicates a high correlation between the current load pattern and the corresponding long-term memory pattern. Simultaneously, predictive performance metrics, such as mean squared error, are calculated for the current prediction model on the independent validation set. and the historical predictive performance metrics stored in this long-term memory pattern. A comparison is performed, and when the feature similarity exceeds a preset recall threshold and When the current prediction performance is lower than the prediction performance of the long-term memory pattern, the system triggers a memory retrieval-fusion event.

[0140] S64. In response to the memory retrieval-fusion event, the parameter snapshot in the long-term memory pattern is used as the teacher model, and the corresponding memory channel in the current prediction model is used as the student model. The knowledge distillation loss is constructed and added as an additional term to the loss function.

[0141] A snapshot of parameters from the long-term memory pattern is loaded into the teacher model, which maintains fixed parameters and is used only for outputting reference prediction results. The corresponding memory channel in the current prediction model is used as the student model, whose parameters are the optimization objectives to be updated. A knowledge distillation loss is constructed, and the Kullback-Leibler Divergence (KL Divergence) is used to measure the difference in the prediction distribution between the student model and the teacher model. ,in, For knowledge distillation loss; The predicted probability distribution output by the teacher model; The predicted probability distribution output by the student model; To predict the time step length; For the teacher model in the first The predicted probability of a time step; For the student model in the first The predicted probability of a time step; To prevent division by zero for small positive numbers, the knowledge distillation loss is added as an additional term to the original loss function, resulting in an updated total loss function. ,in This is the updated total loss function; The original loss function includes the master prediction loss and the channel isolation loss; The distillation loss weight coefficient is used to adjust the influence of knowledge distillation loss on parameter updates. This loss function guides the student model to inherit stable load pattern knowledge from the teacher model while learning new data features, thus avoiding the loss of long-term memory.

[0142] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0143] Based on the same inventive concept, this application also provides a power grid load intelligent prediction and scheduling device for implementing the aforementioned power grid load intelligent prediction and scheduling method. The solution provided by this device is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more power grid load intelligent prediction and scheduling device embodiments provided below can be found in the limitations of the power grid load intelligent prediction and scheduling method described above, and will not be repeated here.

[0144] In one exemplary embodiment, such as Figure 2 As shown, a smart power grid load forecasting and dispatching device is provided, comprising:

[0145] Data module 201 is used to acquire historical load data of the power grid, weather forecast data and event calendar information, and to perform short-time Fourier transform on the historical load data of the power grid based on a sliding window to obtain the spectral feature vector.

[0146] Boundary module 202 is used to calculate the routing attention weights based on the spectral feature vector and the current context information through a pre-trained memory router, evaluate the significance of concept drift based on the spectral feature vector, and generate adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels; the current context information is obtained by encoding weather forecast data and event calendar information.

[0147] The adaptive module 203 is used to construct training samples based on historical load data of the power grid, spectral feature vectors and actual load values, and to train the prediction model based on multiple heterogeneous memory channels based on the training samples, adaptive learning rate multiplier and forget gate threshold based on minimizing the loss function, so as to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

[0148] The prediction module 204 is used to input historical time-series data and future conditional data into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation, so as to obtain the final load point prediction and the prediction uncertainty interval. The historical time-series data consists of historical power grid load data, weather forecast data corresponding to the historical power grid load data, and event calendar information. The future conditional data consists of weather forecast data and event calendar information for the target future period. The multiple heterogeneous memory channels are configured to learn load patterns with different time scales and stability, and the sub-prediction results output by each memory channel are weighted and aggregated using routing attention weights.

[0149] The scheduling module 205 is used to update the scheduling optimization model based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight, and to generate the unit combination and output plan for the target future time period through the updated scheduling optimization model.

[0150] In one embodiment, the data module 201 is further configured to:

[0151] Extract load subsequences from historical load data of the power grid that meet the conditions of ending at the current time and having a preset window size, and apply Hamming windowing to the load subsequences to obtain windowed load subsequences;

[0152] Perform a Fast Fourier Transform on the windowed load subsequence and calculate the square of the spectral magnitude to obtain the time-spectrum diagram of the current sliding window;

[0153] Integrate the energy corresponding to the spectral components with frequencies less than or equal to a preset low-frequency threshold on the time-spectrum graph to obtain the low-frequency energy components;

[0154] Integrate the energy corresponding to the spectral components on the time-spectrum graph whose frequencies are greater than a preset low-frequency threshold and less than a preset high-frequency threshold to obtain the mid-frequency energy component;

[0155] Integrate the energy corresponding to the spectral components with frequencies greater than or equal to a preset high-frequency threshold on the time-spectrum graph to obtain the high-frequency energy components;

[0156] A spectral feature vector is constructed based on low-frequency energy components, mid-frequency energy components, and high-frequency energy components.

[0157] In one embodiment, the boundary module 202 is further configured to:

[0158] The concept drift significance index is calculated based on the ratio of low-frequency to high-frequency energy components in the spectral feature vector, the variation norm of historical routing attention weights, and the historical prediction mean square error of the memory channel. The formula for calculating the concept drift significance index is as follows: ,in, , and These are weighting coefficients. To prevent small positive numbers from being divided by zero, and These are the high-frequency energy components and low-frequency energy components in the spectral eigenvector, respectively. and These are the routing attention weights for the current time and the previous time, respectively. and These are the prediction mean square errors for the third and second memory channels, respectively.

[0159] For the first memory channel configured to learn long-term, slow-changing patterns, the adaptive learning rate multiplier and forgetting gate threshold are calculated based on the concept drift significance index. The formula for calculating the adaptive learning rate multiplier of the first memory channel is as follows: The formula for calculating the forgetting threshold of the first memory channel is: ,in, , , and These are the preset hyperparameters for the first memory channel. This is the significance index for concept drift;

[0160] For the second memory channel configured with a learning day cycle and baseline load pattern, the adaptive learning rate multiplier and forgetting gate threshold of the second memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the second memory channel is as follows: The formula for calculating the forgetting gate threshold of the second memory channel is: ,in and The preset hyperparameters for the second memory channel;

[0161] For the third memory channel configured to learn short-term mutations and events, the adaptive learning rate multiplier and forgetting gate threshold of the third memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the third memory channel is as follows: The formula for calculating the forgetting threshold of the third memory channel is: ,in, , , and These are the preset hyperparameters for the third memory channel.

[0162] In one embodiment, the adaptive module 203 is further configured to:

[0163] The historical load data of the power grid in the training samples is input into the prediction model based on multiple heterogeneous memory channels to obtain the initial sub-prediction results output by each memory channel;

[0164] The mean square error between the weighted aggregation result of each initial sub-prediction result and the true load value in the training sample is calculated to obtain the main prediction loss;

[0165] The real load value is decomposed into corresponding low-frequency, mid-frequency, and high-frequency true values ​​by using a preset bandpass filter.

[0166] The negative correlation coefficients between the initial sub-prediction results of the first memory channel and the low-frequency truth component, the negative correlation coefficients between the initial sub-prediction results of the second memory channel and the mid-frequency truth component, and the negative correlation coefficients between the initial sub-prediction results of the third memory channel and the high-frequency truth component are calculated respectively.

[0167] The channel isolation loss is obtained by multiplying each negative correlation coefficient by a preset isolation weight and then summing the results. The formula for calculating the channel isolation loss is as follows: ,in, The function is the Pearson negative correlation coefficient. , and These are the initial sub-prediction results for the first, second, and third memory channels, respectively. , and The low-frequency true value components, mid-frequency true value components, and high-frequency true value components are identified separately. , and For isolation weights;

[0168] The main prediction loss and channel isolation loss are weighted and summed according to a preset ratio to obtain the loss function, and the gradient of the loss function with respect to the network parameters of each memory channel is calculated through the backpropagation algorithm.

[0169] For each memory channel, the gradient is multiplied by the adaptive learning rate multiplier to obtain the adjusted gradient. The proposed update amount of the network parameters is calculated based on the adjusted gradient. It is then determined whether the ratio of the proposed update amount to the current absolute value of the network parameters exceeds the forget gate threshold to obtain the judgment result.

[0170] If the judgment result is that the proportion exceeds the forget gate threshold, the proposed update amount of the network parameters is trimmed to the range bounded by the forget gate threshold, and the network parameters of the prediction model based on multiple heterogeneous memory channels are updated based on the updated proposed update amount to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

[0171] In one embodiment, the prediction module 204 is further configured to:

[0172] Historical time-series data and future conditional data are input into an adaptive prediction model based on multiple heterogeneous memory channels to obtain the load sub-prediction sequence for each memory channel.

[0173] The load sub-prediction sequences of each memory channel are linearly weighted and aggregated based on routing attention weights to obtain the final load point prediction; the calculation formula for the final load point prediction is as follows: ,in, , and These are the load sub-prediction sequences for the first memory channel, the second memory channel, and the third memory channel, respectively. , and For routing attention weights;

[0174] Based on the historical prediction error statistics of the first, second, and third memory channels, the uncertainty variance of the load sub-prediction sequence of each memory channel is estimated, and the overall uncertainty variance of the final load point prediction is calculated according to the routing attention weight and the uncertainty variance of each memory channel.

[0175] The predicted uncertainty interval is calculated based on the overall uncertainty variance and the preset confidence level parameters.

[0176] In one embodiment, the scheduling module 205 is further configured to:

[0177] Using the final load point forecast as the baseline load curve, and combining it with a load uncertainty set constructed from the forecast uncertainty interval, a robust optimization method is employed to find the worst-case load scenario that maximizes the total system operating cost within the load uncertainty set. The optimization problem that minimizes the objective function while satisfying the constraint set under the worst-case load scenario is then solved, yielding the unit combination and output plan. The objective function is to minimize the total system operating cost, which includes generator fuel cost, unit start-up and shutdown cost, reserve capacity cost, and load deviation penalty cost. The constraint set includes system power balance constraints, upper and lower limits of each generator's output - ramp rate constraints, minimum start-up and shutdown time constraints, network power flow safety constraints, and positive and negative spinning reserve capacity constraints. The positive and negative spinning reserve capacity constraints satisfy... ,in, For positive and negative rotational reserve capacity, Based on positive and negative rotational reserve capacity This is the proportionality coefficient. This refers to the component in the routing attention weights corresponding to the third memory channel.

[0178] In one embodiment, an optimization module is also included, for:

[0179] The norm of change of network parameters of the first memory channel and the second memory channel within the sliding time window is obtained. When the norm of change of the first memory channel or the second memory channel is detected to be continuously lower than the preset stability threshold for more than the preset duration, it is determined that the load mode learned by the corresponding memory channel has entered a stable state and the archiving event is triggered.

[0180] In response to an archive event, a statistical summary of the network parameters, spectral feature vector, and current context information of the memory channel currently in a stable state is encapsulated into a long-term memory pattern.

[0181] Before each training of the prediction model or before the final load point prediction, calculate the spectral feature vector and current context information, and the feature similarity with each long-term memory pattern. When the feature similarity exceeds the preset recall threshold and the prediction performance of the current prediction model on the validation set is lower than the prediction performance of the long-term memory pattern, trigger the memory retrieval-fusion event.

[0182] In response to the memory retrieval-fusion event, the parameter snapshot in the long-term memory pattern is used as the teacher model, and the corresponding memory channel in the current prediction model is used as the student model. The knowledge distillation loss is constructed and added as an additional term to the loss function.

[0183] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0184] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0185] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0186] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent forecasting and scheduling of power grid load, characterized in that, The method includes: Historical load data of the power grid, weather forecast data, and event calendar information are acquired, and a short-time Fourier transform is performed on the historical load data of the power grid based on a sliding window to obtain a spectral feature vector; Based on the spectral feature vector and the current context information, the routing attention weights are calculated through a pre-trained memory router, and the concept drift saliency is evaluated based on the spectral feature vector to generate adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels; the current context information is obtained by encoding the weather forecast data and the event calendar information. Training samples are constructed based on the historical load data of the power grid, the spectral feature vector, and the actual load value. Based on minimizing the loss function, the prediction model based on multiple heterogeneous memory channels is trained using the training samples, the adaptive learning rate multiplier, and the forget gate threshold to obtain an adaptive prediction model based on multiple heterogeneous memory channels. Historical time-series data and future conditional data are input into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation to obtain the final load point prediction and prediction uncertainty interval. The historical time-series data consists of the historical load data of the power grid, the weather forecast data corresponding to the historical load data of the power grid, and the event calendar information. The future conditional data consists of the weather forecast data for the target future time period and the event calendar information. The multiple heterogeneous memory channels are configured to learn load patterns with different time scales and stability, and the sub-prediction results output by each memory channel are weighted and aggregated using the routing attention weight. The scheduling optimization model is updated based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight, and the updated scheduling optimization model is used to generate the unit combination and output plan for the target future time period.

2. The method according to claim 1, characterized in that, The method of performing a short-time Fourier transform on the historical load data of the power grid based on a sliding window to obtain a spectral feature vector includes: Extract a load subsequence from the historical load data of the power grid that meets the conditions of ending at the current time and having a preset window size, and apply a Hamming window to the load subsequence to obtain a windowed load subsequence; Perform a Fast Fourier Transform on the windowed load subsequence and calculate the square of the spectral magnitude to obtain the time-spectrum diagram of the current sliding window; Integrate the energy corresponding to the spectral components whose frequencies are less than or equal to a preset low-frequency threshold on the time-spectrum graph to obtain the low-frequency energy components; Integrate the energy corresponding to the spectral components on the time-spectrum diagram whose frequencies are greater than a preset low-frequency threshold and less than a preset high-frequency threshold to obtain the mid-frequency energy component. Integrate the energy corresponding to the spectral components with frequencies greater than or equal to a preset high-frequency threshold on the time-spectrum diagram to obtain high-frequency energy components; The spectral feature vector is constructed based on the low-frequency energy component, the mid-frequency energy component, and the high-frequency energy component.

3. The method according to claim 2, characterized in that, The process of evaluating the significance of concept drift based on the spectral feature vector and generating adaptive learning rate multipliers and forgetting gate thresholds corresponding to multiple heterogeneous memory channels includes: Based on the ratio of the low-frequency energy component to the high-frequency energy component in the spectral feature vector, the variation norm of the historical routing attention weight, and the historical prediction mean square error of the memory channel, the concept drift significance index is calculated; the formula for calculating the concept drift significance index is as follows: ,in, , and These are weighting coefficients. To prevent small positive numbers from being divided by zero, and These are the high-frequency energy component and the low-frequency energy component in the spectral feature vector, respectively. and These are the routing attention weights for the current time and the previous time, respectively. and These are the prediction mean square errors for the third and second memory channels, respectively. For the first memory channel configured to learn long-term, slow-changing patterns, the adaptive learning rate multiplier and the forgetting gate threshold corresponding to the first memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the first memory channel is as follows: The formula for calculating the forgetting gate threshold of the first memory channel is as follows: ,in, , , and These are the preset hyperparameters for the first memory channel. The concept drift significance index; For the second memory channel configured with a learning day cycle and baseline load mode, the adaptive learning rate multiplier and the forgetting gate threshold of the second memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the second memory channel is as follows: The formula for calculating the forgetting gate threshold of the second memory channel is as follows: ,in and These are the preset hyperparameters for the second memory channel; For the third memory channel configured to learn short-term mutations and events, the adaptive learning rate multiplier and the forgetting gate threshold of the third memory channel are calculated based on the concept drift significance index; the formula for calculating the adaptive learning rate multiplier of the third memory channel is as follows: The formula for calculating the forget gate threshold of the third memory channel is as follows: ,in, , , and These are the preset hyperparameters for the third memory channel.

4. The method according to claim 3, characterized in that, The method of training a prediction model based on multiple heterogeneous memory channels using the training samples, the adaptive learning rate multiplier, and the forgetting gate threshold, based on minimizing the loss function, yields an adaptive prediction model based on multiple heterogeneous memory channels, including: The historical load data of the power grid in the training samples is input into the prediction model based on multiple heterogeneous memory channels to obtain the initial sub-prediction results output by each memory channel; The mean square error between the weighted aggregation result of each initial sub-prediction result and the true load value in the training sample is calculated to obtain the main prediction loss; The real load value is decomposed into corresponding low-frequency, mid-frequency, and high-frequency true value components by a preset bandpass filter. Calculate the negative correlation coefficient between the initial sub-prediction result of the first memory channel and the low-frequency truth component, the negative correlation coefficient between the initial sub-prediction result of the second memory channel and the mid-frequency truth component, and the negative correlation coefficient between the initial sub-prediction result of the third memory channel and the high-frequency truth component, respectively. The channel isolation loss is obtained by multiplying each of the negative correlation coefficients by a preset isolation weight and then summing the results; the formula for calculating the channel isolation loss is as follows: ,in, The function is the Pearson negative correlation coefficient. , and The initial sub-prediction results are respectively for the first memory channel, the second memory channel, and the third memory channel. , and The low-frequency true value components, mid-frequency true value components, and high-frequency true value components are identified separately. , and For isolation weights; The loss function is obtained by weighting and summing the main prediction loss and the channel isolation loss according to a preset ratio, and the gradient of the loss function with respect to the network parameters of each memory channel is calculated by the backpropagation algorithm. For each memory channel, the gradient is multiplied by the adaptive learning rate multiplier to obtain the adjusted gradient, and the proposed update amount of the network parameters is calculated based on the adjusted gradient. It is then determined whether the proportion of the proposed update amount to the current absolute value of the network parameters exceeds the forget gate threshold to obtain the judgment result. If the determination result is that the proportion exceeds the forget gate threshold, then the proposed update amount of the network parameters is trimmed to a range bounded by the forget gate threshold, and the network parameters of the prediction model based on multiple heterogeneous memory channels are updated based on the updated proposed update amount to obtain an adaptive prediction model based on multiple heterogeneous memory channels.

5. The method according to claim 3, characterized in that, The process of inputting historical time-series data and future conditional data into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation, to obtain the final load point prediction and prediction uncertainty interval, includes: The historical time series data and the future conditional data are input into the adaptive prediction model based on multiple heterogeneous memory channels to obtain the load sub-prediction sequence of each memory channel; The load sub-prediction sequences of each memory channel are linearly weighted and aggregated based on the routing attention weights to obtain the final load point prediction; the calculation formula for the final load point prediction is as follows: ,in, , and The load sub-prediction sequences are respectively the first memory channel, the second memory channel, and the third memory channel. , and The route attention weight; Based on the historical prediction error statistics of the first memory channel, the second memory channel, and the third memory channel, the uncertainty variance of the load sub-prediction sequence of each memory channel is estimated respectively, and the overall uncertainty variance of the final load point prediction is calculated according to the routing attention weight and the uncertainty variance of each memory channel. The predicted uncertainty interval is calculated based on the overall uncertainty variance and the preset confidence level parameter.

6. The method according to claim 3, characterized in that, The process of updating the scheduling optimization model based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight, and generating the unit combination and output plan for the target future time period using the updated scheduling optimization model, includes: Using the final load point prediction as the baseline load curve, and combining it with the load uncertainty set constructed from the prediction uncertainty interval, a robust optimization method is employed to find the worst-case load scenario within the load uncertainty set that maximizes the total system operating cost. The optimization problem that satisfies the constraint set and minimizes the objective function under the worst-case load scenario is then solved to obtain the unit combination and the output plan. The objective function is to minimize the total system operating cost, which includes power generation fuel cost, unit start-up and shutdown cost, reserve capacity cost, and load deviation penalty cost. The constraint set includes system power balance constraints, upper and lower limits of each generator unit's output - ramp rate constraints, minimum start-up and shutdown time constraints, network power flow safety constraints, and positive and negative spinning reserve capacity constraints. The positive and negative spinning reserve capacity constraints satisfy... ,in, For positive and negative rotational reserve capacity, Based on positive and negative rotational reserve capacity This is the proportionality coefficient. The component in the routing attention weights that corresponds to the third memory channel.

7. The method according to claim 4, characterized in that, The method further includes: The norm of change of the network parameters of the first memory channel and the second memory channel within the sliding time window is obtained. When the norm of change of the first memory channel or the second memory channel is detected to be continuously lower than the preset stability threshold for more than the preset duration, it is determined that the load mode learned by the corresponding memory channel has entered a stable state and an archiving event is triggered. In response to the archiving event, the network parameter snapshot of the memory channel currently in the stable state, the spectral feature vector, and the statistical summary of the current context information are encapsulated into a long-term memory mode; Before each training of the prediction model or before the final load point prediction, the spectral feature vector and the current context information are calculated, along with the feature similarity of each long-term memory pattern. When the feature similarity exceeds a preset recall threshold and the prediction performance of the current prediction model on the validation set is lower than the prediction performance of the long-term memory pattern, a memory retrieval-fusion event is triggered. In response to the memory retrieval-fusion event, the parameter snapshot in the long-term memory pattern is used as the teacher model, the corresponding memory channel in the current prediction model is used as the student model, a knowledge distillation loss is constructed, and the knowledge distillation loss is added as an additional term to the loss function.

8. A smart power grid load forecasting and dispatching device, characterized in that, The device includes: The data module is used to acquire historical load data of the power grid, weather forecast data and event calendar information, and to perform short-time Fourier transform on the historical load data of the power grid based on a sliding window to obtain a spectral feature vector. The boundary module is used to calculate the routing attention weights based on the spectral feature vector and the current context information through a pre-trained memory router, evaluate the significance of concept drift based on the spectral feature vector, and generate adaptive learning rate multipliers and forget gate thresholds corresponding to multiple heterogeneous memory channels; the current context information is obtained by encoding the weather forecast data and the event calendar information. An adaptive module is used to construct training samples based on the historical load data of the power grid, the spectral feature vector and the actual load value, and to train the prediction model based on multiple heterogeneous memory channels based on the training samples, the adaptive learning rate multiplier and the forget gate threshold based on minimizing the loss function, so as to obtain an adaptive prediction model based on multiple heterogeneous memory channels. The prediction module is used to input historical time-series data and future conditional data into an adaptive prediction model based on multiple heterogeneous memory channels for isolated forward propagation, to obtain the final load point prediction and prediction uncertainty interval. The historical time-series data consists of the historical load data of the power grid, the weather forecast data corresponding to the historical load data of the power grid, and the event calendar information. The future conditional data consists of the weather forecast data for the target future time period and the event calendar information. The multiple heterogeneous memory channels are configured to learn load patterns with different time scales and stability, and the sub-prediction results output by each memory channel are weighted and aggregated using the routing attention weight. The scheduling module is used to update the scheduling optimization model based on the final load point prediction, the prediction uncertainty interval, and the routing attention weight, and to generate the unit combination and output plan for the target future time period through the updated scheduling optimization model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. 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 of any one of claims 1 to 7.

Images