Network computing power load prediction method, device and equipment for power system
By combining LSTM networks and multilayer perceptrons, the load and associated features of the power system are independently modeled and adaptively weighted, solving the coupling interference problem in the mixed modeling of multi-source heterogeneous features and improving the accuracy and robustness of load prediction.
Patent Information
- Application Number
- CN202611211933.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-06
AI Technical Summary
Existing technologies are prone to coupling interference when modeling multi-source heterogeneous features in power systems, and the feature fusion process relies on fixed weights, resulting in a lack of accuracy and robustness in load prediction results.
A pre-set LSTM network is used to analyze the changing trends of historical load sequences. A pre-set multilayer perceptron is used to perform nonlinear mapping analysis on the power correlation feature matrix. By combining the statistics of historical load and power correlation feature matrix, a gating coefficient is generated. The residual calculation formula is used to perform deviation analysis to achieve load target prediction.
By adopting a dual-channel independent modeling architecture, the mutual interference of parameter spaces caused by the direct splicing and mixing of heterogeneous features is avoided, the source of load fluctuation driving is distinguished, and the accuracy and robustness of the prediction results are improved through adaptive weight adjustment.
Smart Images
Figure CN123048988A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of power systems, and in particular to network computing load prediction methods, apparatus and equipment for power systems. Background Technology
[0002] With the rapid development of the digital economy, cloud computing, artificial intelligence, and big data analytics have become widespread, leading to an explosive growth in the scale of data centers, which have become significant high-energy-consuming loads in the power system. The concept of power computing collaborative scheduling has emerged in response. Its core lies in leveraging the spatiotemporal flexibility of computing loads to dynamically adjust the spatiotemporal distribution of computing tasks by sensing power signals such as electricity price fluctuations and renewable energy output, thereby achieving peak shaving and valley filling, and promoting green and low-carbon energy use. In this scenario, computing loads exhibit dual-driven characteristics of endogenous inertia and exogenous scheduling, resulting in more complex evolution patterns. Accurate load forecasting is the core foundation for achieving efficient collaborative scheduling.
[0003] Current computing load forecasting technology primarily focuses on time series modeling. Recurrent neural networks and their variant, Long Short-Term Memory networks, are widely used to uncover the periodicity and trends of load patterns due to their ability to capture long-term temporal dependencies. Transformer models based on self-attention mechanisms further improve prediction accuracy through global correlation modeling. Some studies attempt to incorporate power-related features, using feature concatenation to input multi-source data into a single network to reflect the impact of power-side factors on load.
[0004] However, existing technologies still have obvious shortcomings. First, the direct mixing of multi-source heterogeneous features in modeling results in significant differences in the physical meaning and evolution scale of business time-series features and power features, which can easily lead to coupling interference in the parameter space, and the model cannot accurately distinguish the driving source of load fluctuations. Second, the feature fusion process relies on statically fixed weights, which cannot adaptively adjust the prediction strategy according to the power signal strength, resulting in insufficient robustness in the scenario. Summary of the Invention
[0005] This application provides a network computing load prediction method, apparatus, and equipment for power systems, which solves the technical problems of existing technologies that are difficult to resist coupling interference between features based on multi-source heterogeneous feature hybrid modeling, and feature fusion process relies on fixed weights, resulting in a lack of accuracy and robustness in load prediction results.
[0006] In view of this, the first aspect of this application provides a network computing load prediction method for power systems, including:
[0007] A pre-defined LSTM network is used to analyze the changing trends of historical load sequences to obtain a baseline load prediction value.
[0008] By performing nonlinear mapping analysis on the power correlation feature matrix through a preset multilayer sensor, load adjustment quantities are generated.
[0009] Based on the historical load sequence and the power correlation feature matrix, statistical quantities are calculated, and the calculated statistical vectors are input into a preset gating network to generate gating coefficients.
[0010] By combining the gating coefficient and residual calculation formula, the load adjustment amount is used to perform deviation analysis on the load baseline prediction value to obtain the load target prediction value.
[0011] Preferably, the step of generating load adjustment amounts by performing nonlinear mapping analysis on the power correlation feature matrix using a preset multilayer sensor includes:
[0012] The initial correlation characteristic matrix of the power system is normalized to generate the power correlation characteristic matrix.
[0013] The power correlation feature matrix is reconstructed into a two-dimensional feature matrix along the batch and time dimensions;
[0014] The two-dimensional feature matrix is input into a preset multi-layer perception layer for nonlinear mapping analysis to obtain the load mapping result.
[0015] The load mapping result is reshaped to obtain the load adjustment amount.
[0016] Preferably, the step of calculating statistics based on the historical load sequence and the power correlation feature matrix, and inputting the calculated statistical vector into a preset gating network to generate gating coefficients, includes:
[0017] Statistical calculations are performed on the historical load sequence along the time dimension to obtain the load mean and load standard deviation;
[0018] Statistical calculations are performed on the power correlation feature matrix along the feature dimension to obtain the power mean and power standard deviation.
[0019] The load mean, load standard deviation, power mean, and power standard deviation are stacked to generate a statistical vector;
[0020] The statistical vector is input into a preset gating network to analyze the dynamic impact of power correlation characteristics on the load and calculate the gating coefficient.
[0021] Preferably, the step of calculating statistics based on the historical load sequence and the power correlation feature matrix, and inputting the calculated statistical vector into a preset gating network to generate gating coefficients, further includes:
[0022] Based on the Huber loss function and a preset learning rate, the model parameters of the initial network model are trained offline using the Adam optimizer to obtain an offline optimized model.
[0023] The offline optimization model is fine-tuned based on prediction loss using real-time data samples to obtain an optimized prediction model.
[0024] The optimized prediction model includes the preset LSTM network, the preset multilayer perceptron, and the preset gated network.
[0025] Preferably, the residual calculation formula is expressed as follows:
[0026]
[0027] in, This represents the predicted load target value at time step t. This represents the baseline load forecast. Indicates the gating coefficient. This indicates the load adjustment amount.
[0028] The second aspect of this application provides a network computing power load prediction device for power systems, comprising:
[0029] The baseline prediction unit is used to perform trend analysis on the historical load sequence using a preset LSTM network to obtain the load baseline prediction value.
[0030] The adjustment mapping unit is used to perform nonlinear mapping analysis on the power correlation feature matrix through a preset multilayer sensor to generate load adjustment amount;
[0031] The parameter calculation unit is used to perform statistical calculations based on the historical load sequence and the power correlation feature matrix, and input the calculated statistical vector into a preset gating network to generate gating coefficients.
[0032] The load prediction unit is used to combine the gating coefficient and the residual calculation formula, and use the load adjustment amount to perform deviation analysis on the load baseline prediction value to obtain the load target prediction value.
[0033] Preferably, the adjustment mapping unit is specifically used for:
[0034] The initial correlation characteristic matrix of the power system is normalized to generate the power correlation characteristic matrix.
[0035] The power correlation feature matrix is reconstructed into a two-dimensional feature matrix along the batch and time dimensions;
[0036] The two-dimensional feature matrix is input into a preset multi-layer perception layer for nonlinear mapping analysis to obtain the load mapping result.
[0037] The load mapping result is reshaped to obtain the load adjustment amount.
[0038] Preferably, the parameter calculation unit is specifically used for:
[0039] Statistical calculations are performed on the historical load sequence along the time dimension to obtain the load mean and load standard deviation;
[0040] Statistical calculations are performed on the power correlation feature matrix along the feature dimension to obtain the power mean and power standard deviation.
[0041] The load mean, load standard deviation, power mean, and power standard deviation are stacked to generate a statistical vector;
[0042] The statistical vector is input into a preset gating network to analyze the dynamic impact of power correlation characteristics on the load and calculate the gating coefficient.
[0043] Preferably, it further includes a model optimization unit, specifically used for:
[0044] Based on the Huber loss function and a preset learning rate, the model parameters of the initial network model are trained offline using the Adam optimizer to obtain an offline optimized model.
[0045] The offline optimization model is fine-tuned based on prediction loss using real-time data samples to obtain an optimized prediction model.
[0046] The optimized prediction model includes the preset LSTM network, the preset multilayer perceptron, and the preset gated network.
[0047] A third aspect of this application provides a network computing load prediction device for power systems, the device including a processor and a memory;
[0048] The memory is used to store program code and transmit the program code to the processor;
[0049] The processor is used to execute the network computing load prediction method for power systems as described in the first aspect, according to the instructions in the program code.
[0050] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:
[0051] This application provides a network computing power load prediction method for power systems, including: using a pre-set LSTM network to analyze the changing trends of historical load sequences to obtain a load baseline prediction value; using a pre-set multilayer perceptron to perform nonlinear mapping analysis on the power correlation feature matrix to generate a load adjustment amount; calculating statistics based on the historical load sequence and the power correlation feature matrix, and inputting the calculated statistical vector into a pre-set gating network to generate gating coefficients; combining the gating coefficients and residual calculation formulas, using the load adjustment amount to perform deviation analysis on the load baseline prediction value to obtain a load target prediction value.
[0052] The network-based load forecasting method for power systems provided in this application employs a dual-channel independent modeling architecture. It uses an LSTM network to analyze the time-dimensional load change trends of historical load sequences and a multilayer perceptron to perform nonlinear mapping analysis on the power-related feature matrix. This completes feature extraction analysis of the two types of features in independent parameter spaces, avoiding mutual interference in parameter spaces caused by directly splicing and mixing heterogeneous features. It can distinguish different driving sources of load fluctuations. Furthermore, it calculates statistical vectors based on historical load and power features, and inputs them into a gating network to calculate gating coefficients. These coefficients are used to correct prediction bias. In other words, the fusion weights used in the gating network prediction process are generated from data statistical characteristics, rather than static, fixed constants. These coefficients can adaptively adjust with the power signal strength, thereby ensuring the accuracy and robustness of the prediction results. Therefore, this application solves the technical problems of existing technologies that, based on multi-source heterogeneous feature mixing modeling, are difficult to resist coupling interference between features, and that the feature fusion process relies on fixed weights, resulting in a lack of accuracy and robustness in load forecasting results. Attached Figure Description
[0053] Figure 1 A flowchart illustrating the network computing power load prediction method for power systems provided in this application embodiment;
[0054] Figure 2 This is a schematic diagram of the structure of a network computing load prediction device for power systems provided in an embodiment of this application. Detailed Implementation
[0055] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.
[0056] For easier understanding, please refer to Figure 1The embodiments of the network computing power load prediction method for power systems provided in this application include:
[0057] Step 101: Use a preset LSTM network to analyze the changing trends of historical load sequences and obtain the load baseline prediction value.
[0058] It should be noted that Long Short-Term Memory (LSTM) is a special type of recurrent neural network specifically designed to solve the gradient vanishing / exploding problem when traditional RNNs process long sequences. Its core is to achieve controllable retention and updating of information through a gating mechanism. It introduces cell states as a "memory conveyor belt" that runs through the sequence, using addition operations to transfer information, avoiding exponential gradient decay, and achieving effective retention of long-term dependent information.
[0059] The preset LSTM network in this embodiment is a pre-trained and optimized network model that can analyze the periodic trend changes of historical load sequences in the time dimension. The network uses gating mechanisms and cell states to capture high-order time dependencies in long sequences and learn regular patterns such as daily and weekly cycles. Based on this network, the time dependencies driven by business inertia can be captured to obtain the load baseline prediction value.
[0060] Specifically, the pre-defined LSTM network can be viewed as a load analysis channel composed of stacked long short-term memory networks and fully connected mapping layers, with Dropout regularization introduced between layers to suppress overfitting. This channel takes a historical load sequence L of shape [B,T,1] as input, where B is the batch size and T is the historical window length. After the load channel encodes the sequence L step-by-step, the hidden state vector h of the last time step of the last layer is extracted. last The load baseline prediction value is mapped to a shape of [B,H] by a fully connected layer. Where H is the prediction time step, reflecting the trend component that can be predicted solely by the historical load's own evolution pattern, i.e., the prediction result of load change trend under the influence of business.
[0061] Step 102: Perform nonlinear mapping analysis on the power correlation feature matrix using a preset multilayer sensor to generate load adjustment.
[0062] It should be noted that the power correlation feature matrix is composed of multi-dimensional feature data related to the power system and computing load, specifically including but not limited to power market prices, market price changes, new energy power generation, system energy power generation changes, weekday date sine coding and cosine coding.
[0063] A multilayer perceptron (MLP) is a feedforward artificial neural network composed of multiple neurons arranged hierarchically, including an input layer, hidden layers, and an output layer. Neurons in different layers are connected by weights, and information propagates sequentially from the input layer to the output layer without feedback connections. This embodiment uses a pre-defined MLP to perform nonlinear mapping analysis on the power correlation feature matrix to generate load adjustment amounts. Furthermore, the power signal mapping analysis is performed synchronously with the load baseline prediction analysis described above, simultaneously achieving load prediction and quantifying load deviations caused by power-side factors, providing adjustment amounts for subsequent prediction deviation adjustments.
[0064] Further, step 102 includes:
[0065] The initial correlation characteristic matrix of the power system is normalized to generate the power correlation characteristic matrix.
[0066] The power correlation feature matrix is reconstructed into a two-dimensional feature matrix along the batch and time dimensions;
[0067] The two-dimensional feature matrix is input into a preset multi-layer perception layer for nonlinear mapping analysis to obtain the load mapping result;
[0068] The load mapping results are reshaped to obtain the load adjustment amount.
[0069] It should be noted that the initial correlation characteristic matrix in this embodiment includes heterogeneous power system characteristic information of multiple dimensions. Therefore, in order to avoid the impact of inconsistent dimensions on subsequent calculation and analysis, this embodiment will perform normalization processing on the matrix composed of these data information in advance to eliminate the risk caused by inconsistent dimensions. Therefore, the power correlation characteristic matrix used in subsequent calculation and analysis is the normalized power correlation characteristic matrix.
[0070] Specifically, in this embodiment, the power channel takes a future multidimensional power correlation feature matrix P of shape [B,H,D] as input, where D=6, including six dimensions: power market price, market price change, renewable energy generation, renewable energy generation change, weekday date sine code, and cosine code. First, the power correlation feature matrix P needs to be reshaped into a two-dimensional feature matrix of [B×H,D] along the dimensions of batch B and time H. This matrix is then input into a pre-defined multilayer perceptron with shared weights for time-step nonlinear mapping analysis. Finally, the output is reshaped back into a two-dimensional shape of [B,H] to obtain the power correlation load adjustment Δ. power This load adjustment represents the deviation in load increase or decrease caused by fluctuations in electricity prices and changes in renewable energy output. Furthermore, the point-by-point mapping mechanism in this embodiment, which pre-sets weights shared across all time steps of the multilayer sensor, reduces the number of parameters from H independent sets to a single set of shared parameters, thus lowering model complexity and the risk of overfitting.
[0071] Step 103: Calculate statistics based on historical load sequences and power correlation feature matrices, and input the calculated statistical vectors into a preset gating network to generate gating coefficients.
[0072] It should be noted that the historical load sequence and the power correlation feature matrix are both the original load data sequence and the power feature normalized data matrix. The corresponding statistics for each can be calculated separately, and then the statistics are input into a preset gating network for dynamic weight analysis to generate gating coefficients. It can be found that the gating coefficients are affected by the dynamic changes of the power correlation feature matrix, thereby achieving adaptive dynamic adjustment. Compared with the traditional static fixed weights, the adaptive dynamic coefficients in this embodiment are more in line with the application characteristics of the actual scenario, more resistant to the impact of power scenario fluctuations, and can improve the accuracy and robustness of subsequent prediction results.
[0073] The statistics in this embodiment specifically include the mean and standard deviation. Moreover, the calculation of the statistics is independent of the gating network. In other words, the global statistics do not depend on the convergence state of the model parameters. The gating network can be provided with stable input early in the model training process, which decouples the weight update of the gating network from the convergence progress of the load and power dual channels, reducing the interference problems caused by the coupling of heterogeneous features.
[0074] Further, step 103 includes:
[0075] Statistical calculations were performed on the historical load series along the time dimension to obtain the load mean and load standard deviation;
[0076] Statistical calculations are performed on the power correlation feature matrix along the feature dimensions to obtain the power mean and power standard deviation.
[0077] The load mean, load standard deviation, power mean, and power standard deviation are stacked to generate a statistical vector;
[0078] The statistical vector is input into a preset gating network to analyze the dynamic impact of power correlation characteristics on the load and calculate the gating coefficient.
[0079] It should be noted that the statistics of the historical load sequence are calculated along the time dimension, specifically including the load mean and load standard deviation; while the statistics of the power correlation feature matrix are calculated along the feature dimension, specifically including the power mean and power standard deviation. By stacking the four global statistics of load mean, load standard deviation, power mean, and power standard deviation in one dimension, a statistical vector of shape [B,4] can be obtained.
[0080] By inputting the statistical vector into a pre-defined gating network for adaptive weight analysis, the gating coefficient can be obtained. Adaptive weight analysis determines the dynamic impact of power correlation characteristics on the load. When the statistical characteristics of the power features indicate that electricity price fluctuations and changes in renewable energy output have a significant impact on the load, the gating coefficient approaches 1, and the adjustment of the power channel is fully adopted, improving prediction accuracy. When the power correlation signal is weak, the gating coefficient approaches 0, and the prediction value degenerates into a pure time-series baseline prediction, avoiding the introduction of additional errors by low-quality power features and ensuring output stability in extreme scenarios.
[0081] Furthermore, the preset gating network in this embodiment consists of a fully connected layer, ReLU, another fully connected layer, and a sigmoid activation function in sequence. The input is a statistical vector G, and the output is the gating coefficients. Furthermore, the input global statistical vector does not depend on the convergence state of the model parameters, decoupling the adaptive coefficient update of the gating network from the convergence progress of the load and power channels, ensuring that the adaptive update is not affected.
[0082] Furthermore, step 103, preceding the following, also includes:
[0083] Based on the Huber loss function and a preset learning rate, the model parameters of the initial network model are trained offline using the Adam optimizer to obtain an offline optimized model.
[0084] The offline optimization model is fine-tuned based on prediction loss using real-time data samples to obtain an optimized prediction model.
[0085] The optimized prediction model includes the preset LSTM network, the preset multilayer perceptron, and the preset gated network.
[0086] Because existing technologies generally lack online continuous optimization mechanisms, models relying solely on offline training struggle to cope with data distribution drift, resulting in a continuous decline in accuracy over long-term operation and an inability to adapt to dynamically changing power-computing power collaborative scheduling scenarios. Therefore, this embodiment designs a model optimization training approach combining offline and online methods: an offline supervised training phase and an online fine-tuning phase; both phases involve load and power data. In the offline training phase, the training set can be divided into a front-end training subset and a back-end validation subset according to chronological order; global statistics can be calculated based solely on the training subset, avoiding data leakage caused by overlapping sliding windows; then, training data loaders and validation data loaders are constructed based on the normalized data.
[0087] In the specific offline training phase, all batches of the training loader can be traversed in training mode to enable the initial network model to perform forward prediction operations, including dual-channel analysis, gating coefficient calculation, and prediction error between the real-time predicted load value and the actual load value. This error is calculated based on the Huber loss function, so it actually calculates the Huber loss between the predicted value and the actual value. Then, the prediction error is backpropagated through the Adam optimizer to update all model parameters.
[0088] In the offline validation phase, predictions are made by inputting a subset of the validation data into the model through the validation loader, and the validation loss, mean absolute error, and root mean square error are calculated. Based on this, the training effect of the model is evaluated, and the optimal offline training model, i.e., the offline optimized model, is selected. Furthermore, the preset learning rate used in the model training phase of this embodiment is determined based on the ReduceLROnPlateau scheduler, which uses a minimization mode for decay. If the validation loss does not decrease within a preset number of consecutive rounds, the current learning rate is automatically multiplied by a decay factor, the value of which ranges from [0,1]. Additionally, if the current validation loss is lower than the historical best validation loss, the current model parameters are saved as the optimal model parameters, and the early stopping counter is reset; otherwise, the counter continues to increment. If the early stopping counter reaches the preset maximum number of rounds, training ends prematurely.
[0089] The online fine-tuning phase essentially uses the trained offline optimized model as the initial model for the online phase, sets the online learning rate (typically 1 / 30 to 1 / 10 of the offline learning rate), and initializes an online normalizer based on the Welford incremental algorithm. Statistics can be updated using only the current value, historical mean, historical second central moment, and cumulative sample count, eliminating the need to store all historical data and keeping memory overhead constant.
[0090] Real-time data samples are new samples acquired sequentially from the real-time data stream. They also contain several historical load sequences and feature matrices relating them to future power consumption. Furthermore, to calculate subsequent prediction losses, corresponding actual load values are also acquired. Next, forward prediction analysis can be performed based on the initial model to obtain online predicted values. Then, the Huber loss (prediction loss) between the online predicted values and the actual values is calculated. The model parameters are then updated step-by-step via backpropagation using the Adam optimizer with an online learning rate, resulting in an optimized prediction model. This optimized prediction model includes a pre-set LSTM network, a pre-set multilayer perceptron, and a pre-set gated network. This fine-tuning process allows the model parameters to adapt to changes in data distribution in a continuous and gradual manner.
[0091] In addition, to suppress catastrophic forgetting, this embodiment can also set up an experience replay buffer of fixed capacity and use a circular queue structure to cache historical samples. Whenever the number of online steps is an integer multiple of the preset replay interval, historical samples are randomly sampled from the buffer and mixed with the current samples to form a mini-batch, and an additional forward propagation and backward update are performed to suppress catastrophic forgetting in continuous online learning.
[0092] Moreover, it can continuously calculate the mean absolute error, root mean square error, and exponential moving average of the error within the sliding window. If the indicators continuously deteriorate beyond the preset threshold, it will trigger weight rollback, restore the model parameters to the checkpoint where the accuracy verification was passed most recently, and further reduce the online learning rate.
[0093] It should be noted that, regardless of whether it is the offline optimization model obtained from offline training or the optimized prediction model obtained from online fine-tuning, the output is a load prediction sequence in normalized space. Therefore, it is necessary to use the statistical vectors obtained during the training phase, namely the load mean and load standard deviation, to perform inverse normalization on the load prediction sequence and restore it to the predicted value at the true load scale, that is, the load target prediction value.
[0094] Step 104: Combining the gating coefficient and residual calculation formula, the load adjustment amount is used to perform deviation analysis on the load baseline prediction value to obtain the load target prediction value.
[0095] Furthermore, the residual calculation formula is expressed as:
[0096]
[0097] in, This represents the predicted load target value at time step t. This represents the baseline load forecast. Indicates the gating coefficient. This represents the load adjustment amount. The residual correction calculation can be performed at each prediction time step. The output results from both the power and load channels are fused for correction analysis to obtain a normalized spatial prediction sequence, the normalized load prediction sequence. By performing inverse normalization on this sequence based on the global statistics mentioned above, the true load scale, i.e., the predicted load target value, can be obtained.
[0098] It should be noted that the use of dynamic adaptive gating coefficients in the prediction bias correction process in this embodiment can improve the prediction accuracy, while the use of residual calculation formulas allows the power channel to learn only the bias components without fitting the complete load level, which greatly reduces the learning dimension and training rounds of the power channel.
[0099] The network computing power load prediction method for power systems provided in this application adopts a dual-channel independent modeling architecture. It uses an LSTM network to analyze the time-dimensional load change trend of historical load sequences and a multilayer perceptron to perform nonlinear mapping analysis on the power-related feature matrix. This completes feature extraction analysis of the two types of features in independent parameter spaces, avoiding mutual interference in parameter spaces caused by direct splicing and mixing of heterogeneous features, and can distinguish different driving sources of load fluctuations. Furthermore, it calculates statistical vectors based on historical load and power features, and inputs them into a gating network to calculate gating coefficients. Based on these coefficients, prediction bias correction is achieved. In other words, the fusion weights used in the gating network prediction process are generated from data statistical characteristics, rather than static, fixed constants. These coefficients can adaptively adjust with the power signal strength, thereby ensuring the accuracy and robustness of the prediction results. Therefore, this application can solve the technical problems of existing technologies that, based on multi-source heterogeneous feature mixing modeling, are difficult to resist coupling interference between features, and that the feature fusion process relies on fixed weights, resulting in a lack of accuracy and robustness in load prediction results.
[0100] For easier understanding, please refer to Figure 2 This application provides an embodiment of a network computing power load prediction device for power systems, including:
[0101] The baseline prediction unit 201 is used to perform trend analysis on the historical load sequence using a preset LSTM network to obtain the load baseline prediction value.
[0102] The adjustment mapping unit 202 is used to perform nonlinear mapping analysis on the power correlation feature matrix through a preset multilayer sensor to generate load adjustment amount;
[0103] The parameter calculation unit 203 is used to perform statistical calculations based on historical load sequences and power correlation feature matrices, and input the calculated statistical vectors into a preset gating network to generate gating coefficients.
[0104] The load prediction unit 204 is used to combine the gating coefficient and the residual calculation formula, and use the load adjustment amount to perform deviation analysis on the load baseline prediction value to obtain the load target prediction value.
[0105] Furthermore, the mapping unit 202 is adjusted, specifically for:
[0106] The initial correlation characteristic matrix of the power system is normalized to generate the power correlation characteristic matrix.
[0107] The power correlation feature matrix is reconstructed into a two-dimensional feature matrix along the batch and time dimensions;
[0108] The two-dimensional feature matrix is input into a preset multi-layer perception layer for nonlinear mapping analysis to obtain the load mapping result;
[0109] The load mapping results are reshaped to obtain the load adjustment amount.
[0110] Furthermore, the parameter calculation unit 203 is specifically used for:
[0111] Statistical calculations were performed on the historical load series along the time dimension to obtain the load mean and load standard deviation;
[0112] Statistical calculations are performed on the power correlation feature matrix along the feature dimensions to obtain the power mean and power standard deviation.
[0113] The load mean, load standard deviation, power mean, and power standard deviation are stacked to generate a statistical vector;
[0114] The statistical vector is input into a preset gating network to analyze the dynamic impact of power correlation characteristics on the load and calculate the gating coefficient.
[0115] Furthermore, it also includes a model optimization unit 205, specifically used for:
[0116] Based on the Huber loss function and a preset learning rate, the model parameters of the initial network model are trained offline using the Adam optimizer to obtain an offline optimized model.
[0117] The offline optimization model is fine-tuned based on prediction loss using real-time data samples to obtain an optimized prediction model.
[0118] The optimized prediction model includes the preset LSTM network, the preset multilayer perceptron, and the preset gated network.
[0119] This application also provides a network computing load prediction device for power systems, the device including a processor and memory;
[0120] The memory is used to store program code and transfer the program code to the processor;
[0121] The processor is used to execute the network computing load prediction method for power systems in the above method embodiments according to the instructions in the program code.
[0122] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0124] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of this application through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.
[0126] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A network computing power load forecasting method for power systems, characterized in that, include: A pre-defined LSTM network is used to analyze the changing trends of historical load sequences to obtain a baseline load prediction value. By performing nonlinear mapping analysis on the power correlation feature matrix through a preset multilayer sensor, load adjustment quantities are generated. Based on the historical load sequence and the power correlation feature matrix, statistical quantities are calculated, and the calculated statistical vectors are input into a preset gating network to generate gating coefficients. By combining the gating coefficient and residual calculation formula, the load adjustment amount is used to perform deviation analysis on the load baseline prediction value to obtain the load target prediction value.
2. The network computing power load prediction method for power systems according to claim 1, characterized in that, The step of generating load adjustment amounts by performing nonlinear mapping analysis on the power correlation feature matrix using a preset multilayer perceptron includes: The initial correlation characteristic matrix of the power system is normalized to generate the power correlation characteristic matrix. The power correlation feature matrix is reconstructed into a two-dimensional feature matrix along the batch and time dimensions; The two-dimensional feature matrix is input into a preset multi-layer perception layer for nonlinear mapping analysis to obtain the load mapping result. The load mapping result is reshaped to obtain the load adjustment amount.
3. The network computing power load prediction method for power systems according to claim 1, characterized in that, The step of calculating statistics based on the historical load sequence and the power correlation feature matrix, and inputting the calculated statistical vector into a preset gating network to generate gating coefficients includes: Statistical calculations are performed on the historical load sequence along the time dimension to obtain the load mean and load standard deviation; Statistical calculations are performed on the power correlation feature matrix along the feature dimension to obtain the power mean and power standard deviation. The load mean, load standard deviation, power mean, and power standard deviation are stacked to generate a statistical vector; The statistical vector is input into a preset gating network to analyze the dynamic impact of power correlation characteristics on the load and calculate the gating coefficient.
4. The network computing power load prediction method for power systems according to claim 1, characterized in that, The step of calculating statistics based on the historical load sequence and the power correlation feature matrix, and inputting the calculated statistical vector into a preset gating network to generate gating coefficients, also includes: Based on the Huber loss function and a preset learning rate, the model parameters of the initial network model are trained offline using the Adam optimizer to obtain an offline optimized model. The offline optimization model is fine-tuned based on prediction loss using real-time data samples to obtain an optimized prediction model. The optimized prediction model includes the preset LSTM network, the preset multilayer perceptron, and the preset gated network.
5. The network computing power load prediction method for power systems according to claim 1, characterized in that, The residual calculation formula is expressed as follows: in, This represents the predicted load target value at time step t. This represents the baseline load forecast. Indicates the gating coefficient. This indicates the load adjustment amount.
6. A network computing power load forecasting device for power systems, characterized in that, include: The baseline prediction unit is used to perform trend analysis on the historical load sequence using a preset LSTM network to obtain the load baseline prediction value. The adjustment mapping unit is used to perform nonlinear mapping analysis on the power correlation feature matrix through a preset multilayer sensor to generate load adjustment amount; The parameter calculation unit is used to perform statistical calculations based on the historical load sequence and the power correlation feature matrix, and input the calculated statistical vector into a preset gating network to generate gating coefficients. The load prediction unit is used to combine the gating coefficient and the residual calculation formula, and use the load adjustment amount to perform deviation analysis on the load baseline prediction value to obtain the load target prediction value.
7. The network computing power load prediction device for power systems according to claim 6, characterized in that, The adjustment mapping unit is specifically used for: The initial correlation characteristic matrix of the power system is normalized to generate the power correlation characteristic matrix. The power correlation feature matrix is reconstructed into a two-dimensional feature matrix along the batch and time dimensions; The two-dimensional feature matrix is input into a preset multi-layer perception layer for nonlinear mapping analysis to obtain the load mapping result. The load mapping result is reshaped to obtain the load adjustment amount.
8. The network computing power load prediction device for power systems according to claim 6, characterized in that, The parameter calculation unit is specifically used for: Statistical calculations are performed on the historical load sequence along the time dimension to obtain the load mean and load standard deviation; Statistical calculations are performed on the power correlation feature matrix along the feature dimension to obtain the power mean and power standard deviation. The load mean, load standard deviation, power mean, and power standard deviation are stacked to generate a statistical vector; The statistical vector is input into a preset gating network to analyze the dynamic impact of power correlation characteristics on the load and calculate the gating coefficient.
9. The network computing power load prediction device for power systems according to claim 6, characterized in that, It also includes a model optimization unit, specifically used for: Based on the Huber loss function and a preset learning rate, the model parameters of the initial network model are trained offline using the Adam optimizer to obtain an offline optimized model. The offline optimization model is fine-tuned based on prediction loss using real-time data samples to obtain an optimized prediction model. The optimized prediction model includes the preset LSTM network, the preset multilayer perceptron, and the preset gated network.
10. A network computing power load forecasting device for power systems, characterized in that, The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the network computing load prediction method for power systems according to any one of the claims 1-5, based on the instructions in the program code.