A source-load multitask collaborative prediction method based on parallel time sequence PLE

Abstract

The application discloses a source-load multi-task collaborative prediction method based on parallel time sequence PLE, first, historical meteorological data, wind power data and load data are acquired, and after pretreatment and variational mode decomposition, a multi-source heterogeneous feature matrix is constructed; then, a progressive layered extraction multi-task network is constructed, and task-specific expert networks and cross-task shared expert networks are set for wind power and load tasks; each task adopts a double-channel prediction tower with parallel Informer and BiLSTM channels to extract global long-term trend and local instantaneous fluctuation characteristics respectively; an extreme value weighted loss is introduced in the loss function to enhance the perception of peak nodes, and an asymmetric Pinball loss based on quantile regression is used to guide conservative wind power prediction and load surplus prediction; finally, local weighted linear regression is used in the output layer to dynamically extract the wind speed-power physical boundary to hard-cut the wind power prediction value. The application effectively suppresses multi-task negative transfer and asymmetric risk, and improves the prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of source-load prediction technology, and specifically relates to a source-load multi-task collaborative prediction method based on parallel temporal PLE. Background Technology

[0002] As the "dual carbon" goals are being implemented more deeply, the proportion of new energy sources, represented by wind power, in scenarios such as direct green electricity supply is constantly increasing. However, the volatility and intermittency of wind power output and the uncertainty of load demand pose serious challenges to the real-time balance and dispatch security of the power system. Accurate wind power and load forecasting are key prerequisites for ensuring the safe, economical, and stable operation of direct green electricity supply systems.

[0003] In existing technologies, mainstream prediction methods mostly employ symmetric loss functions (such as mean squared error, MSE) to train models. These loss functions treat "positive bias" and "negative bias" equally during optimization, meaning overestimation and underestimation incur the same penalty. However, in practical engineering scenarios such as direct green energy supply, the impact of prediction bias in different directions on system safety is significantly asymmetric: overestimating wind power output (predicted values ​​higher than actual values) may lead to excessive wind power reserves in the scheduling plan, resulting in unnecessary wind curtailment and resource waste; underestimating load demand (predicted values ​​lower than actual values) may lead to insufficient generation capacity, triggering load shedding or even power outages. Existing technologies lack an asymmetric risk perception mechanism that can proactively guide the model to generate "conservative predictions" or "safety margin predictions," resulting in prediction results lacking necessary safety constraints in practical engineering.

[0004] Furthermore, existing models such as Multi-Scale Convolutional Neural Networks (MFCNN) and Long Short-Term Memory Networks (LSTM) learn mapping patterns solely from the statistical distribution of historical data. These methods do not embed hard physical boundary conditions such as wind turbine cut-in / cut-out wind speeds and rated power into the learning framework. In areas with drastic weather fluctuations or extremely low power output, data noise can easily cause residual fluctuations in the model that violate physical principles. Moreover, traditional methods often model wind power and load demand in isolation, neglecting the complex spatiotemporal complementarity and dynamic interaction coupling mechanism between source-side wind power and load-side demand mediated by the meteorological system. Even when some multi-task learning frameworks attempt to jointly predict source and load, their sharing mechanisms struggle to effectively separate shared and unique features between tasks with high heterogeneity, easily leading to a "seesaw effect" (performance improvement in one task while decline in another) and negative transfer phenomena.

[0005] Finally, existing models exhibit weak predictive capabilities and bias control issues under critical extreme scenarios. Grid dispatch is extremely sensitive to peak load times and periods of severe wind power fluctuations. Traditional models tend to fit the average distribution of data samples to minimize global residuals, resulting in a significant decrease in sensitivity when dealing with extreme peak data exhibiting strong non-stationarity. The predicted trajectory tends to shrink towards the mean, failing to accurately capture instantaneous spikes or precipitates, further exacerbating the risks of violating physical logic and deviating from safety margins. Summary of the Invention

[0006] The purpose of this invention is to provide a source-load multi-task collaborative prediction method based on parallel temporal PLE, which solves the problems of poor security, low prediction accuracy and output that violate physical common sense caused by the symmetry loss function, lack of physical boundaries and weak perception of extreme scenes in the prior art.

[0007] To achieve the above objectives, this invention provides a source-load multi-task collaborative prediction method based on parallel temporal PLE, comprising the following steps: S1. Obtain historical meteorological data, historical power load data, and historical wind power data for the target area, and perform data preprocessing; S2. Variational mode decomposition technology is used to decompose the preprocessed wind power sequence and load sequence to obtain a series of intrinsic mode functions (IMFs) with independent frequency characteristics. The decomposed mode components are then fused with meteorological driving factors and time signature features to construct a multi-source heterogeneous feature matrix. S3. Construct a multi-task learning framework based on progressive hierarchical extraction. The multi-task learning framework sets up two task-specific expert networks for wind power forecasting and load forecasting tasks respectively to represent their unique fluctuation patterns. At the same time, it sets up four cross-task shared expert networks to uniformly capture the daily / weekly periodicity and seasonal trends shared by the source and load. The outputs of the task-specific expert networks and the shared expert networks are adaptively weighted and routed through a multi-level gating network. S4. Construct a parallel dual-path prediction tower for each prediction task. The prediction tower includes an Informer path and a BiLSTM path. The Informer path uses a probabilistic sparse self-attention mechanism and temporal feature distillation technology to extract the global long-term periodic patterns and macroscopic evolution trends corresponding to low-frequency modal features. The BiLSTM path captures the microscopic local instantaneous fluctuations and spike jump patterns corresponding to high-frequency modal features along the forward and reverse temporal dimensions. The feature matrices output by the two paths are fused and mapped through the linear projection layer of the prediction tower to obtain the source load prediction values ​​for future time steps. S5. Design an extreme value loss function with a differential weighting mechanism to increase attention to critical scheduling nodes of load peaks or sudden wind power changes, assign a weight coefficient greater than 1 to samples identified as extreme nodes, and assign a weight of 1 to the remaining samples. S6. Design a Pinball loss function based on quantile regression to quantify and penalize symmetric errors, and actively guide the model to generate strategic directional bias in prediction. For source-end wind power tasks, set a quantile below 0.5 to guide the model to adopt a conservative estimation principle; for load-end load tasks, set a quantile above 0.5 to guide the model to adopt a surplus estimation principle. S7. Weight the mean squared error loss, extreme value weighted loss and asymmetric risk perception loss to construct the overall loss function, which is used to train the multi-task learning framework. S8. Based on local weighted linear regression, the wind speed-power physical boundary is dynamically extracted, and the wind power prediction value output by the prediction tower is subject to boundary truncation constraint according to the corresponding wind speed in the current prediction period, so that the final wind power prediction result strictly conforms to the physical engineering boundary of wind turbine operation.

[0008] Furthermore, in step S3, the multi-level gating network adopts an adaptive gating mechanism based on Softmax. In the j-layer extraction network, the multi-level gating network of task k performs weighted routing of each expert output through the adaptive Softmax gating mechanism, and the output expression is:

[0009] in, It is a task by The weight matrix is ​​used as input; It is a task In the The selection matrix is ​​extracted from the network.

[0010] Furthermore, in step S4, the feature sequence input to the prediction tower is X, and the Informer pathway calculates the attention of the input feature sequence X as follows:

[0011] in, For query matrix; The key matrix; For the reason before A sparse query matrix consisting of contribution queries; It is a value matrix; Let K be the dimension of the key matrix; () is a normalization function used to map the attention weights to the [0,1] interval. Only the first few weights are selected in the calculation. Each query participates in the attention operation, reducing the computational complexity from Down to ; The computational complexity is that of traditional attention mechanisms; represents the computational complexity of the probabilistic sparse attention mechanism; L is the length of the input feature sequence. The intermediate hidden states output by the encoder in the Informer path are further fed into the decoder for cross-attention calculation with the target concatenated sequence to obtain the full-time domain decoding matrix. To extract the global representation of the future-oriented prediction window, the prediction length of the decoding matrix at the end of the temporal dimension is truncated. The submatrix on the matrix yields the task-specific global trend feature matrix. :

[0012] in, This represents the total timing length of the decoding matrix; This indicates that the time series dimension is truncated using a range index; The efficient computation method for capturing data along both the forward and reverse temporal dimensions of the BiLSTM path is as follows:

[0013] in, This represents the input feature vector of the unified input feature X at time t; Let be the hidden state of the forward LSTM at time t. This represents the hidden state of the inverse LSTM at time t. The final hidden state at time t is obtained by concatenating the forward and reverse hidden states; After forward propagation through a multi-layer bidirectional recurrent network stack, the BiLSTM path outputs a full-time state sequence. By extracting the full-time state sequence Predicting length at the end of the time series dimension Submatrices on top are used to achieve dimension alignment, thereby extracting local fluctuation feature matrices. :

[0014] in, The total time length of the input sequence; Finally, the global trend feature matrix output by the Informer path is... Local fluctuation characteristic matrix of BiLSTM channel output Tensor concatenation along the feature dimension yields the fused feature matrix. Finally, the fused feature matrix is ​​fed into the output linear projection layer of the prediction tower for mapping, outputting single-step non-iterative source load prediction data for future time steps.

[0015] Furthermore, in step S5, the formula for calculating the extreme value loss function is as follows:

[0016] Where n is the total number of sample nodes; i is the index of the sample node; This represents the true value of the i-th node. Let be the predicted value of the i-th node; Let be the differential weight coefficient of the i-th node; when When identified as an extreme node, the weight coefficient Other regular nodes .

[0017] Furthermore, in step S6, the formula for calculating the Pinball loss function is as follows:

[0018] in, For source-end wind power tasks, set quantiles. For load-side tasks, set quantiles. .

[0019] Furthermore, in step S7, the formula for calculating the overall loss function is:

[0020] in, This is the mean square error loss of wind power. This is the load mean square error loss; , These represent extreme value losses for wind power and load, respectively. , Pinball losses for wind power and load, respectively; , The weight balancing coefficients are selected through iterative search using an automated hyperparameter framework. , .

[0021] Furthermore, step S8 includes the following steps: S8.1 For the normalized wind speed-power time series data, outlier data points are removed, and equal sampling is performed. Within the obtained equal sample set, the wind speed range is divided into multiple equidistant sub-intervals. The maximum power point and minimum power point are extracted in each sub-interval, and the original upper boundary candidate point set and lower boundary candidate point set are constructed respectively. S8.2. A quadratic polynomial locally weighted linear regression model is used to perform point-by-point sliding fitting on the candidate point set, dynamically generating initial upper and lower boundary curves that continuously change with wind speed. The fitting model expression is:

[0022] in, The normalized boundary fitting value for wind power; The normalized target wind speed query point; , and These represent the local constant term, first-order coefficient, and second-order coefficient of the locally weighted linear regression model at the current wind speed point, respectively, used to characterize the local nonlinearity of the power curve; for any given wind speed query point... its neighboring candidate points Gaussian kernel weights The calculation formula is:

[0023] in, This refers to the Gaussian kernel bandwidth parameter; S8.3. The initial boundary curve is physically normalized and corrected by combining the wind turbine's cut-in velocity, cut-out velocity, and rated power. The formula is:

[0024] in, and These are the upper and lower boundary values ​​of the dynamic wind power after correction according to physical rules; and These are the initial upper and lower boundary fitted values ​​output by the locally weighted linear regression model, respectively. This is the normalized unit power constant; S8.4 At the output terminal of the forward propagation of the prediction tower, based on the extracted dynamic physical boundary, a safety margin is dynamically assigned according to the wind speed. The dynamic upper and lower boundaries are scaled proportionally using an amplification factor of 1.01 and a reduction factor of 0.99, respectively, to form a more flexible dynamic hard constraint threshold. The corresponding flexible upper and lower boundaries are dynamically obtained according to the current wind speed, and the original wind power prediction value is forcibly truncated to the range of these boundaries.

[0025] Furthermore, in step S8.1, the equalization sampling includes: setting a normalized wind speed threshold V. th=0.2 is used as the boundary between low wind speed area and medium-high wind speed area. If the number of samples detected in medium-high wind speed area is greater than 100, all samples in that area are completely retained. In low wind speed area, 200 samples are drawn by random sampling without replacement and mixed with medium-high wind speed samples to obtain a balanced sample set. Within the obtained balanced sample set, the wind speed range is divided into 200 equidistant sub-intervals.

[0026] Furthermore, in step S1, the historical meteorological data includes seven key meteorological features, including wind direction, wind speed, temperature, relative humidity, surface air pressure, total precipitation, and total surface irradiance. The historical meteorological data is preprocessed by extracting periodic attributes and normalizing. The historical power load data and historical wind power data are preprocessed by using the isolated forest algorithm to identify nonlinear outliers, using the Z-score-based statistical method to detect abnormal samples, filling in missing values, and normalizing.

[0027] Furthermore, in step S2, the optimal number of decomposition modes K for variational mode decomposition is 8, and the penalty factor α is 2000.

[0028] After adopting the above solution, the beneficial effects of the present invention are as follows: 1. To address the challenge of collaborative forecasting modeling between two ends caused by the heterogeneity of power sources and loads in existing power systems, this invention constructs a progressive hierarchical extraction multi-task learning framework (PLE framework). This framework establishes independent task-specific expert networks for wind power forecasting and load forecasting tasks, while simultaneously setting up a cross-task shared expert network. Furthermore, it utilizes a multi-level gating network for adaptive weighted routing, achieving dynamic separation of common and unique features between the source and load ends. This mechanism overcomes the "seesaw effect" and negative migration interference commonly found in traditional parameter sharing mechanisms, significantly improving the overall statistical accuracy of joint forecasting between the two ends.

[0029] 2. To address the asymmetric risks in power system operation, namely "overestimating wind power leading to curtailment and underestimating load leading to power outages," this invention abandons the traditional symmetric loss function and introduces Pinball loss based on quantile regression into the source-load multi-task prediction framework. With only a slight sacrifice in marginal global numerical fitting accuracy, it proactively and strategically guides the model's prediction error towards a specific bias direction that aligns with the actual dispatch safety of the power grid. By setting a low quantile (e.g., 0.1) for wind power tasks, the model is guided to adopt a conservative estimation principle, significantly mitigating the risks of wind curtailment, power abandonment, and system tripping caused by overestimating wind power output. Conversely, by setting a high quantile (e.g., 0.9) for load tasks, the model is guided to adopt a surplus estimation principle, avoiding power supply gaps and grid outages caused by underestimating electricity demand. Through this asymmetric directional bias design, this invention transforms uncontrollable symmetric random statistical errors into a controllable safety margin for power system operation.

[0030] 3. This invention designs an extreme value loss function with a differential weighting mechanism, increasing the sensitivity to high-risk operating nodes such as peak loads and sudden wind power fluctuations, and increasing the weight of loss penalties. This forces the model to break through the limitations of traditional global residual minimization and enhances its sensitivity to non-stationary extreme data. This method effectively suppresses the defect of predicted trajectories shrinking towards the mean, accurately capturing instantaneous peaks and sharp drops, and providing reliable support for grid dispatch safety under extreme conditions.

[0031] 4. This invention embeds the hard physical boundaries of wind turbine operation (cut-in wind speed, cut-out wind speed, rated power) into the end of the prediction model as dynamic constraints. It adaptively extracts the upper and lower boundary curves of wind speed-power from historical data using locally weighted linear regression (LWLR), and uses these curves, corrected by physical rules, as hard cut-off conditions for forward propagation. This technology completely eliminates irrational and singular mutation values ​​that violate engineering physics logic from the source, significantly improving the physical reliability and engineering applicability of the prediction results.

[0032] 5. This invention integrates Informer and BiLSTM dual-pathways in parallel within a task-specific prediction tower. Informer utilizes a probabilistic sparse self-attention mechanism to efficiently extract global long-term periodic patterns and macroscopic evolutionary trends, reducing computational complexity from O(L²) to O(Lln L). BiLSTM captures high-frequency local instantaneous fluctuations and spike jumps along bidirectional time series. The fusion of features from both pathways enables the model to grasp slow changes such as daily / weekly seasonality and accurately respond to rapid processes such as sharp rises / falls in power within minutes, comprehensively improving prediction accuracy.

[0033] 6. The prediction method of this invention can not only break through the limitations of traditional information silo-based independent prediction and achieve deep integration of high statistical accuracy and physical consistency of power grid operation, but also provide methodological support with high physical credibility and core data decision input with high reliability for day-ahead and intraday robust optimization scheduling and risk margin control management of new power systems. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a wind power correlation analysis diagram of the present invention; Figure 3 This is a load correlation analysis diagram of the present invention; Figure 4 This is a VMD exploded view of wind power output according to the present invention; Figure 5 This is a load VMD decomposition diagram of the present invention; Figure 6 This is a diagram of the PLE-Informer-BiLSTM source-load joint prediction architecture of the present invention; Figure 7 This is a flowchart of the dynamic boundary extraction process of the present invention; Figure 8 This is a comparison chart of the normalized wind power prediction results and error indices of various models in this invention; Figure 9 This is a comparison chart of the normalized load prediction results and error indices of the various models in this invention. Figure 9 Legend and Figure 8 Same, using Figure 8 Legend of the diagram. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Key references Figures 1-9 This invention provides a source-load multi-task collaborative prediction method based on parallel temporal PLE, comprising the following steps: S1. Obtain hourly historical meteorological data, historical power load data, and historical wind power data for the target area within a complete annual cycle, and perform data preprocessing.

[0037] The historical meteorological data includes seven key meteorological features, including wind direction, wind speed, temperature, relative humidity, surface air pressure, total precipitation, and total surface irradiance. The preprocessing of the historical meteorological data includes: performing sine and cosine coding to extract periodic attributes, and using the Z-score normalization algorithm to normalize all meteorological features and unify the variable scale to eliminate the influence of dimensions.

[0038] The preprocessing of historical power load data and historical wind power data includes: for highly volatile wind power data, the Isolation Forest algorithm is used to identify nonlinear outliers; for load sequences with significant periodicity, a Z-score-based statistical method is used to detect outliers; for missing values, a hybrid strategy consisting of forward imputation, backward imputation, and median interpolation is used to complete the data; historical lag features and rolling statistical features are constructed for the source load data, and the Z-score normalization algorithm is used to normalize the source load data and the constructed features.

[0039] After preprocessing, correlation analysis was performed on the meteorological data and source load data (analysis figure shown). Figure 2 and Figure 3As shown in the figure, key meteorological driving features are screened, and redundant or irrelevant variables are eliminated to provide a basis for subsequent feature construction and fusion.

[0040] S2. Variational Mode Decomposition (VMD) technology is introduced, with the optimal number of decomposition modes K=8 and the penalty factor α=2000. The preprocessed non-stationary wind power and load sequences are decomposed separately to obtain a series of intrinsic mode functions (IMFs) with independent frequency characteristics. The decomposed mode components are then fused with meteorological driving factors and time-stamped features to construct a multi-source heterogeneous feature matrix. The decomposition diagram is shown below. Figure 4 and Figure 5 As shown.

[0041] S3. Construct a multi-task learning framework based on progressive hierarchical extraction, namely the PLE framework. This framework sets up two task-specific expert networks for wind power forecasting and load forecasting tasks respectively to represent their unique fluctuation patterns, and simultaneously sets up four cross-task shared expert networks to uniformly capture the daily / weekly periodicity and seasonal trends shared by source and load. An adaptive weighted routing of the outputs of the task-specific and shared expert networks is performed through a multi-level gating network. Specifically, a Softmax-based adaptive gating mechanism is used for weighted routing. In the j-layer extraction network, the multi-level gating network for task k performs weighted routing of each expert output through an adaptive Softmax gating mechanism. The output expression is:

[0042] in, It is a task by The weight matrix is ​​used as input; It is a task In the The selection matrix is ​​extracted from the network.

[0043] S4. The network terminal adopts a task-specific prediction tower architecture, constructing a parallel dual-path prediction tower for each prediction task. The prediction tower includes an Informer path and a BiLSTM (Bidirectional Long Short-Term Memory) path. That is, this invention adopts a PLE-Informer-BiLSTM source-load joint prediction architecture, as shown in the diagram below. Figure 6 As shown.

[0044] The Informer pathway utilizes a probabilistic sparse self-attention mechanism and temporal feature distillation to specifically extract the global long-term periodic patterns and macroscopic evolutionary trends corresponding to low-frequency modal features. The feature sequence X input to the prediction tower is then used by the Informer pathway to calculate the attention of the input feature sequence X.

[0045] in, For query matrix; The key matrix; For the reason before A sparse query matrix consisting of high-contribution queries; A value matrix; Let K be the dimension of the key matrix; () is a normalization function used to map the attention weights to the [0,1] interval. Only the first few weights are selected in the calculation. Each query participates in the attention operation, reducing the computational complexity from Down to ; The computational complexity is that of traditional attention mechanisms; represents the computational complexity of the probabilistic sparse attention mechanism; L represents the length of the input feature sequence (i.e., the number of historical time steps).

[0046] The intermediate hidden states output by the encoder in the Informer path are further fed into the decoder for cross-attention calculation with the target concatenated sequence to obtain the full-time domain decoding matrix. To extract the global representation of the future-oriented prediction window, the prediction length of the decoding matrix at the end of the temporal dimension is truncated. The submatrix on the matrix yields the task-specific global trend feature matrix. :

[0047] in, This represents the total timing length of the decoding matrix; This indicates that the time series dimension is truncated using a range index; The BiLSTM pathway utilizes a bidirectional recurrent neural network structure to efficiently capture the microscopic local instantaneous fluctuations and spike transition patterns corresponding to high-frequency modal features along both forward and reverse time dimensions. The calculation method is as follows:

[0048] in, This represents the input feature vector of the unified input feature X at time t; Let be the hidden state of the forward LSTM at time t. This represents the hidden state of the inverse LSTM at time t. The final hidden state at time t is obtained by concatenating the forward and reverse hidden states; After forward propagation through a multi-layer bidirectional recurrent network stack, the BiLSTM path outputs a full-time state sequence. Similarly, by extracting the full-time state sequence Predicting length at the end of the time series dimension Submatrices on top are used to achieve dimension alignment, thereby extracting local fluctuation feature matrices. :

[0049] in, The total time length of the input sequence; Finally, the global trend feature matrix output by the Informer path is... Local fluctuation characteristic matrix of BiLSTM channel output Tensor concatenation along the feature dimension yields the fused feature matrix. Finally, the fused feature matrix is ​​fed into the output linear projection layer of the prediction tower for mapping, and the output is single-step non-iterative source load prediction data for future time steps. The input sliding window is set to 96 steps, and the decoder starts the token with a length of 48 steps.

[0050] S5. Design an extreme value loss function with a differential weighting mechanism to increase attention to critical scheduling nodes during load peaks or sudden wind power changes. The calculation formula is as follows:

[0051] Where n is the total number of sample nodes; i is the index of the sample node; This represents the true value of the i-th node. Let be the predicted value of the i-th node; Let be the differential weight coefficient of the i-th node, when When identified as an extreme node, the weight coefficient (The wind power task uses a sliding window to compare neighborhood points in real time to extract local extreme points; the load task uses a percentile threshold strategy to define the top 90% of samples as high-load extreme nodes), and the remaining regular nodes. .

[0052] S6. Design a Pinball loss (adaptive asymmetric bias loss) function based on quantile regression to quantify and penalize symmetric errors, actively guiding the model's predictions to generate a strategic directional bias. The calculation formula is as follows:

[0053] in, Based on the quantile values, differentiated quantile-based current distribution configurations are implemented according to the power grid engineering safety requirements at both the source and load ends. For wind power tasks at the source end, a lower quantile is set (below 0.5, preferred). This guides the model to adopt a conservative estimation principle, significantly reducing the risk of wind curtailment and turbine termination caused by overestimating power output; for load-side tasks, a higher quantile (above 0.5, preferred) is set. This forces the model to favor grid margin (i.e., surplus estimation principle) to avoid power supply gaps and power outages caused by underestimating load.

[0054] S7. The mean squared error loss (MSE, used to characterize the basic regression error between the predicted and the true values), the extreme value weighted loss, and the asymmetric risk perception loss are weighted and summed to construct the overall loss function, which is used to train the multi-task learning framework. The formula for calculating the overall loss function is as follows:

[0055] in, This is the mean square error loss of wind power. This is the load mean square error loss; , These represent extreme value losses for wind power and load, respectively. , Pinball losses for wind power and load, respectively; , To determine the optimal co-weighting balance coefficients, an iterative search using the hyperparameter automation framework (Optuna) is employed. , .

[0056] S8. Based on locally weighted linear regression (LWLR), the wind speed-power physical boundary is dynamically extracted, and the wind power prediction value output by the prediction tower is subject to boundary truncation constraint according to the corresponding wind speed of the current prediction period, so that the final wind power prediction result strictly conforms to the physical engineering boundary of wind turbine operation; such as Figure 6 As shown, the specific steps include: S8.1 For the normalized wind speed-power time series data, outlier data points with normalized wind speeds outside the range [0,1] are removed, and balanced sampling is performed. Balanced sampling includes: setting a normalized wind speed threshold Vth=0.2 (taking the extreme point 1 of the wind speed range corresponding to a wind speed of 35m / s as an example, a threshold of 0.2 corresponds to a wind speed of 7m / s) as the boundary between low-wind-speed and medium-high-wind-speed regions. That is, wind speeds less than or equal to 7m / s are considered medium-high-wind-speed regions, and wind speeds greater than 7m / s are considered low-wind-speed regions. If the number of samples detected in a medium-high-wind-speed region is greater than 100, all samples in that region are retained. In the low-wind-speed region, a random sampling mechanism without replacement is used to extract 200 samples and mix them with the medium-high-wind-speed samples to balance the data distribution across the entire wind speed range, obtaining a balanced sample set. This avoids the overwhelming dominance of low-wind-speed samples in the subsequent loss function, ensuring the accuracy and robustness of the boundary extraction across the entire range.

[0057] Within the obtained balanced sample set, the wind speed range is divided into 200 equidistant sub-intervals. The maximum power point and minimum power point are extracted in each sub-interval to construct the original upper boundary candidate point set and lower boundary candidate point set, respectively. The pure boundary feature points are generated by deduplication.

[0058] S8.2. Establish the feature matrix of candidate points in each interval, and use a quadratic polynomial locally weighted linear regression model (LWLR model) to perform point-by-point sliding fitting on the candidate point set, dynamically generating initial upper and lower boundary curves that continuously change with wind speed. The fitting model expression is:

[0059] in, The normalized boundary fitting value for wind power; The normalized target wind speed query point; , and These represent the local constant term, first-order coefficient, and second-order coefficient of the LWLR model at the current wind speed point, respectively, used to characterize the local nonlinearity of the power curve; for any given wind speed query point... its neighboring candidate points Gaussian kernel weights The calculation formula is:

[0060] in, The Gaussian kernel bandwidth parameter is preferably set to [value missing]. =0.05; S8.3 The initial upper and lower boundary curves generated after fitting the LWLR model need to be standardized and corrected in conjunction with the wind turbine's cut-in wind speed, cut-out wind speed, and rated power to ensure that they strictly conform to the actual power engineering. The formula is:

[0061] in, and These are the upper and lower boundary values ​​of the dynamic wind power after correction according to physical rules; and These are the initial upper and lower boundary fitted values ​​output by the locally weighted linear regression model, respectively. The normalized unit power constant is preferably set to 1.0, thereby outputting a dynamic upper and lower boundary that is continuous, smooth, and has rigid physical boundary protection across the entire wind speed range.

[0062] S8.4 At the output terminal of the forward propagation of the prediction tower, based on the extracted dynamic physical boundary, a safety margin is dynamically assigned according to the wind speed. The dynamic upper and lower boundaries are scaled proportionally using an amplification factor of 1.01 and a reduction factor of 0.99, respectively, to form a more flexible dynamic hard constraint threshold. The original wind power prediction value is standardized and forcibly constrained according to the corresponding wind speed of the current prediction period: if the original wind power prediction value exceeds the elastic upper boundary under the corresponding wind speed, it is forcibly truncated to the elastic upper boundary value; if it is lower than the elastic lower boundary under the corresponding wind speed, it is forcibly truncated to the elastic lower boundary value; thereby completely eliminating irrational prediction mutation points that violate the physical principles of wind turbine power generation from the source.

[0063] To verify the effectiveness of the prediction method proposed in this invention, this embodiment uses hourly multidimensional meteorological elements and actual operation data of new source loads in a coastal city in Fujian Province, China in 2021 as the empirical analysis object. Multiple ablation experiments and comparative experiments were designed, and the performance was compared with that of a single-task model.

[0064] Dataset: Hourly resolution data collected over a full year includes seven meteorological features: wind direction, wind speed, temperature, relative humidity, surface air pressure, total precipitation, and total surface irradiance, as well as synchronized wind power and load active power sequences. The dataset is divided into training, validation, and test sets in a 7:1:2 ratio.

[0065] Evaluation metrics: MAE (mean absolute error), RMSE (root mean square error), Bias-Rate, and Peak Bias-Rate are used as quantitative evaluation metrics for prediction accuracy.

[0066] Comparison Model Setup: To verify the contributions of each key module (extreme value weighted loss, asymmetric biased loss, wind power physics hard constraints, and multi-task collaborative mechanism) in the model proposed in this invention, the following 7 sets of comparison models were constructed:

[0067] Figure 8 and Figure 9 The prediction results and error indices of each comparative model on the test set are shown. Table 1 below gives the MAE, RMSE, Bias-Rate, and PeakBias-Rate values ​​of each model on wind power forecasting and load forecasting tasks.

[0068] Table 1 - Comparison of Predictive Performance Indicators of the Model in Wind Power and Load Forecasting Scenarios

[0069] The data above shows that the proposed model M0 significantly outperforms all ablation models and single-task models in terms of wind power and load forecasting accuracy. Compared to M4, M0's wind power MAE is 3.32% lower and its load MAE is 0.012% lower, verifying the overall effectiveness of extreme value loss, asymmetric bias loss, and physical hard constraints. Compared to the single-task model (SW / SL), M0's wind power MAE is 0.202% lower than SW and its load MAE is 0.119% lower than SL, proving that the multi-task PLE framework successfully achieves positive migration between source and load.

[0070] Furthermore, removing the extreme value loss M1 leads to an increase in wind power RMSE, mainly in the peak range, while the increase in MAE is relatively small, indicating that the extreme value loss module mainly suppresses extreme errors. Removing the asymmetric bias loss M2 results in non-conservative errors (positive deviations) in wind power tasks and a tendency to underestimate load tasks. Although the accuracy error of the load decreases slightly, the impact of this bias on grid dispatch security is far greater than the numerical accuracy itself, which is precisely the asymmetric risk problem that this invention focuses on solving. Removing the physical hard constraints of wind power M3 resulted in a small number of outliers in wind power tasks (caused by power mutation points), verifying the necessity of physical boundary embedding to eliminate irrational outputs.

[0071] In summary, the proposed model M0 demonstrates the best performance in terms of statistical accuracy, physical consistency, and scheduling safety adaptability, thus verifying the synergistic effectiveness of the various innovative modules in this invention.

[0072] The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0073] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A source-load multi-task collaborative prediction method based on parallel temporal PLE, characterized in that, Includes the following steps: S1. Obtain historical meteorological data, historical power load data, and historical wind power data for the target area, and perform data preprocessing; S2. Variational mode decomposition technology is used to decompose the preprocessed wind power sequence and load sequence to obtain a series of intrinsic mode functions (IMFs) with independent frequency characteristics. The decomposed mode components are then fused with meteorological driving factors and time signature features to construct a multi-source heterogeneous feature matrix. S3. Construct a multi-task learning framework based on progressive hierarchical extraction. The multi-task learning framework sets up two task-specific expert networks for wind power forecasting and load forecasting tasks respectively to represent their unique fluctuation patterns. At the same time, it sets up four cross-task shared expert networks to uniformly capture the daily / weekly periodicity and seasonal trends shared by the source and load. The outputs of the task-specific expert networks and the shared expert networks are adaptively weighted and routed through a multi-level gating network. S4. Construct a parallel dual-path prediction tower for each prediction task. The prediction tower includes an Informer path and a BiLSTM path. The Informer path uses a probabilistic sparse self-attention mechanism and temporal feature distillation technology to extract the global long-term periodic patterns and macroscopic evolution trends corresponding to low-frequency modal features. The BiLSTM path captures the microscopic local instantaneous fluctuations and spike jump patterns corresponding to high-frequency modal features along the forward and reverse temporal dimensions. The feature matrices output by the two paths are fused and mapped through the linear projection layer of the prediction tower to obtain the source load prediction values ​​for future time steps. S5. Design an extreme value loss function with a differential weighting mechanism to increase attention to critical scheduling nodes of load peaks or sudden wind power changes, assign a weight coefficient greater than 1 to samples identified as extreme nodes, and assign a weight of 1 to the remaining samples. S6. Design a Pinball loss function based on quantile regression to quantify and penalize symmetric errors, and actively guide the model to generate strategic directional bias in prediction. For source-end wind power tasks, set a quantile below 0.5 to guide the model to adopt a conservative estimation principle; for load-end load tasks, set a quantile above 0.5 to guide the model to adopt a surplus estimation principle. S7. Weight the mean squared error loss, extreme value weighted loss and asymmetric risk perception loss to construct the overall loss function, which is used to train the multi-task learning framework. S8. Based on local weighted linear regression, the wind speed-power physical boundary is dynamically extracted, and the wind power prediction value output by the prediction tower is subject to boundary truncation constraint according to the corresponding wind speed in the current prediction period, so that the final wind power prediction result strictly conforms to the physical engineering boundary of wind turbine operation.

2. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 1, characterized in that: In step S3, the multi-level gating network adopts an adaptive gating mechanism based on Softmax. In the j-layer extraction network, the multi-level gating network of task k performs weighted routing of each expert output through the adaptive Softmax gating mechanism, and the output expression is: in, It is a task by The weight matrix is ​​used as input; It is a task In the The selection matrix is ​​extracted from the network.

3. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 2, characterized in that: In step S4, the feature sequence input to the prediction tower is X, and the Informer pathway calculates the attention of the input feature sequence X as follows: in, For query matrix; The key matrix; For the reason before A sparse query matrix consisting of contribution queries; It is a value matrix; Let K be the dimension of the key matrix; () is a normalization function used to map the attention weights to the [0,1] interval. Only the first few weights are selected in the calculation. Each query participates in the attention operation, reducing the computational complexity from Down to ; The computational complexity is that of traditional attention mechanisms; represents the computational complexity of the probabilistic sparse attention mechanism; L is the length of the input feature sequence. The intermediate hidden states output by the encoder in the Informer path are further fed into the decoder for cross-attention calculation with the target concatenated sequence to obtain the full-time domain decoding matrix. To extract the global representation of the future-oriented prediction window, the prediction length of the decoding matrix at the end of the temporal dimension is truncated. The submatrix on the matrix yields the task-specific global trend feature matrix. : in, This represents the total timing length of the decoding matrix; This indicates that the time series dimension is truncated using a range index; The efficient computation method for capturing data along both the forward and reverse temporal dimensions of the BiLSTM path is as follows: in, This represents the input feature vector of the unified input feature X at time t; Let be the hidden state of the forward LSTM at time t. This represents the hidden state of the inverse LSTM at time t. The final hidden state at time t is obtained by concatenating the forward and reverse hidden states; After forward propagation through a multi-layer bidirectional recurrent network stack, the BiLSTM path outputs a full-time state sequence. By extracting the full-time state sequence Predicting length at the end of the time series dimension Submatrices on top are used to achieve dimension alignment, thereby extracting local fluctuation feature matrices. : in, The total time length of the input sequence; Finally, the global trend feature matrix output by the Informer path is... Local fluctuation characteristic matrix of BiLSTM channel output Tensor concatenation along the feature dimension yields the fused feature matrix. Finally, the fused feature matrix is ​​fed into the output linear projection layer of the prediction tower for mapping, outputting single-step non-iterative source load prediction data for future time steps.

4. A source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 1 or 3, characterized in that: In step S5, the formula for calculating the extreme value loss function is: Where n is the total number of sample nodes; i is the index of the sample node; This represents the true value of the i-th node. Let be the predicted value of the i-th node; Let be the differential weight coefficient of the i-th node; when When identified as an extreme node, the weight coefficient Other regular nodes .

5. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 4, characterized in that: In step S6, the formula for calculating the Pinball loss function is: in, For source-end wind power tasks, set quantiles. For load-side tasks, set quantiles. .

6. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 5, characterized in that: In step S7, the formula for calculating the overall loss function is: in, This is the mean square error loss of wind power. This is the load mean square error loss; , These represent extreme value losses for wind power and load, respectively. , Pinball losses for wind power and load, respectively; , The weight balancing coefficients are selected through iterative search using an automated hyperparameter framework. , .

7. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 1, characterized in that: Step S8 includes the following steps: S8.1 For the normalized wind speed-power time series data, outlier data points are removed, and equal sampling is performed. Within the obtained equal sample set, the wind speed range is divided into multiple equidistant sub-intervals. The maximum power point and minimum power point are extracted in each sub-interval, and the original upper boundary candidate point set and lower boundary candidate point set are constructed respectively. S8.

2. A quadratic polynomial locally weighted linear regression model is used to perform point-by-point sliding fitting on the candidate point set, dynamically generating initial upper and lower boundary curves that continuously change with wind speed. The fitting model expression is: in, The normalized boundary fitting value for wind power; The normalized target wind speed query point; , and These represent the local constant term, first-order coefficient, and second-order coefficient of the locally weighted linear regression model at the current wind speed point, respectively, used to characterize the local nonlinearity of the power curve; for any given wind speed query point... its neighboring candidate points Gaussian kernel weights The calculation formula is: in, This refers to the Gaussian kernel bandwidth parameter; S8.

3. The initial boundary curve is physically normalized and corrected by combining the wind turbine's cut-in velocity, cut-out velocity, and rated power. The formula is: in, and These are the upper and lower boundary values ​​of the dynamic wind power after correction according to physical rules; and These are the initial upper and lower boundary fitted values ​​output by the locally weighted linear regression model, respectively. This is the normalized unit power constant; S8.4 At the output terminal of the forward propagation of the prediction tower, based on the extracted dynamic physical boundary, a safety margin is dynamically assigned according to the wind speed. The dynamic upper and lower boundaries are scaled proportionally using an amplification factor of 1.01 and a reduction factor of 0.99, respectively, to form a more flexible dynamic hard constraint threshold. The corresponding flexible upper and lower boundaries are dynamically obtained according to the current wind speed, and the original wind power prediction value is forcibly truncated to the range of these boundaries.

8. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 7, characterized in that: In step S8.1, equalization sampling includes: setting a normalized wind speed threshold V. th =0.2 is used as the boundary between low wind speed area and medium-high wind speed area. If the number of samples detected in medium-high wind speed area is greater than 100, all samples in that area are completely retained. In low wind speed area, 200 samples are drawn by random sampling without replacement and mixed with medium-high wind speed samples to obtain a balanced sample set. Within the obtained balanced sample set, the wind speed range is divided into 200 equidistant sub-intervals.

9. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 1, characterized in that: In step S1, the historical meteorological data includes seven key meteorological features, including wind direction, wind speed, temperature, relative humidity, surface air pressure, total precipitation and total surface irradiance. The historical meteorological data is preprocessed by extracting periodic attributes and normalizing. For historical power load data and historical wind power data, we perform preprocessing including: using the isolated forest algorithm to identify nonlinear outliers, using a Z-score-based statistical method to detect abnormal samples, filling in missing values, and normalization.

10. The source-load multi-task collaborative prediction method based on parallel temporal PLE as described in claim 1, characterized in that: In step S2, the optimal number of decomposition modes in variational mode decomposition is K=8, and the penalty factor is α=2000.

Images