Method for dynamic generation of power market operation risk list for regional climate characteristics

Abstract

The present application relates to power market risk list generation, in particular to a method for dynamically generating a power market operation risk list for regional climate characteristics, which comprises: inputting current typhoon climate characteristics and typhoon abnormal load sequence into a time sequence feature capturing model based on CNN and BiLSTM architecture to generate climate load time correlation features; inputting the climate load time correlation features into a generator based on a convolutional neural network architecture to generate a load sample sequence; inputting the load sample sequence and the climate load time correlation features into a double-branch discriminator to generate load sample authenticity and predicted load, and generating a power market typhoon emergency state operation risk list based on the predicted load; wherein the generator and the double-branch discriminator constitute a CGAN structure. The present application realizes power load prediction under the condition of adapting to the scarcity of typhoon samples, so as to generate a more accurate power market operation risk list and improve the disaster prevention resilience of the power market.

Description

Technical Field

[0001] This invention relates to the generation of electricity market risk lists, and more particularly to a method for dynamically generating electricity market operation risk lists based on regional climate characteristics. Background Technology

[0002] The safe and stable operation of the electricity market is crucial for ensuring regional energy supply and maintaining socio-economic order. Its operational status is highly correlated with regional climate characteristics. Typhoons, as a frequent extreme weather disaster in my country's southeastern coastal region, especially Hainan, significantly disrupt the electricity market. Typhoons are characterized by unpredictable paths, intense intensity, and high destructive power. The accompanying strong winds and torrential rains upon landfall can cause abnormal fluctuations in electricity load. For example, industrial load decreases due to factory shutdowns, while emergency load increases due to flood control and disaster relief efforts. This disrupts the power system's source-load balance, resulting in a mismatch between the total active power generated by power sources and the total active power consumed by all loads. This triggers a series of electricity market operational risks, including drastic price fluctuations, power supply gaps, transaction defaults, and system frequency security imbalances, posing significant challenges to the electricity market's emergency response.

[0003] Therefore, based on the requirement to enhance the disaster resilience of the power market, it is necessary to embed an automatic operation risk list generation method based on conventional power trading management into the power market trading platform. The auxiliary operation risk list helps the platform to more accurately trigger emergency mechanisms such as early warning, flow restriction, and temporary intervention, and to automatically intercept abnormal bids that match different typhoon weather anomalies. This avoids a one-size-fits-all approach to intercepting abnormal bids under extreme weather conditions, adapts to the real-time risk control needs of the spot power market, and ultimately achieves a risk monitoring indicator system that fits the characteristics of the Hainan power market and meets the actual risk control needs of Hainan power trading institutions.

[0004] Currently, traditional risk list generation methods are mostly based on simple statistical correlation between historical typhoon data and power market operation data, failing to fully capture the complex temporal dependence between typhoon climate characteristics and power load. The characteristics of typhoons, such as wind speed, air pressure, and path, change dynamically over time, and their impact on power load has obvious lag and nonlinearity. Traditional methods cannot effectively mine this dynamic correlation feature, resulting in low accuracy and poor timeliness of the generated risk list, making it difficult to predict sudden risks caused by typhoons in advance.

[0005] On the other hand, existing deep learning-based risk prediction methods use single time series models such as LSTM and BiLSTM to predict loads and then generate risk lists based on the prediction results. However, when these methods are applied to typhoon climate load prediction scenarios, there is a lack of historical typhoon samples, and different typhoons have significant differences in intensity and path. Single time series models are difficult to adapt to such small sample and highly volatile scenarios, and are prone to prediction bias.

[0006] Therefore, how to achieve power load forecasting that adapts to the climate characteristics of typhoon areas and the scarcity of typhoon samples, in order to generate a more accurate list of power market operation risks and enhance the disaster resilience of the power market, is a technical problem that needs to be solved. Summary of the Invention

[0007] To address this, the present invention provides a method for dynamically generating a power market operation risk list based on regional climate characteristics. This method overcomes the problems in existing technologies where power load forecasting cannot adapt to the climate characteristics of typhoon regions and where typhoon samples are scarce. As a result, the power market operation risk list is not accurate enough, and the power market has poor disaster resilience.

[0008] To achieve the above objectives, this invention proposes a method for dynamically generating a power market operation risk inventory tailored to regional climate characteristics, comprising: The current typhoon climate feature sequence from the meteorological website and the typhoon abnormal load sequence from the power market trading platform are obtained. The current typhoon climate features and the typhoon abnormal load sequence are then used to generate climate load time correlation features through a time-series feature capture model based on CNN and BiLSTM architecture. The temporal correlation features of the climate load are used to generate a load sample sequence through a generator based on a convolutional neural network architecture; The load sample sequence and climate load temporal correlation features are processed through a dual-branch discriminator to generate load sample accuracy and predicted load, and a list of power market typhoon emergency operation risks is generated based on the predicted load. The generator and the dual-branch discriminator constitute a CGAN structure. The temporal feature capture model and the generator are trained using the wind speed load physical constraint loss function, and the dual-branch discriminator is trained based on the WGAN-GP loss function.

[0009] Furthermore, the process of generating temporal correlation features of climate load includes: The spliced ​​vector of the current typhoon climate feature sequence and the typhoon abnormal load sequence is passed through a preprocessing layer to generate a spliced ​​input vector; The concatenated input vector is passed through a BiLSTM layer to generate a long-term feature sequence of climate load. The long-term climate load feature sequence is used to generate enhanced time-step long-term climate load features through a typhoon wind speed bias self-attention mechanism. The concatenated input vector is passed through a convolutional layer to generate short-term climate load features; The enhanced time-step climate load long-term feature sequence and climate load short-term feature are concatenated along the channel dimension to generate the climate load time-related feature; The temporal feature capture model includes a preprocessing layer, a BiLSTM layer, a self-attention mechanism, and a convolutional layer.

[0010] Furthermore, the processes that generate enhanced long-term characteristics of time-step climate loads include: The time step elements of the long-term characteristic sequence of the climate load are respectively processed through a self-attention mechanism to generate initial time step attention weights; The indicator function based on the current typhoon wind speed and the typhoon effect threshold is multiplied by a bias coefficient to adjust the initial time step attention weights and generate prior bias-enhanced time step attention weights. The prior bias-enhanced time-step attention weights are used to weight and sum the time-step elements of the long-term climate load feature sequence to generate attention-enhanced features. The attention-enhanced features and time-step elements are concatenated into vectors to generate the enhanced time-step climate load long-term features.

[0011] Furthermore, the process of generating the load sample sequence includes: The concatenated vector of the climate load time-related features and random noise is passed through a fully connected reshaping layer to generate an initial feature map; The initial feature map and the climate load time-related features are passed through a conditional batch normalization unit to generate normalized features; The normalized features are passed through a multi-scale gated convolutional layer to generate multi-scale typhoon load abrupt change features; After residual concatenation of the multi-scale typhoon load abrupt change features and the initial feature map, the load sample sequence is generated through a convolutional output layer. The generator includes a fully connected reshaping layer, a conditional batch normalization unit, a multi-scale gated convolutional layer, a residual connection layer, and a convolutional output layer.

[0012] Furthermore, the process of generating normalized features includes: The time-related features of the climate load are used through a multilayer perceptron to generate a typhoon climate intensity scaling factor and a typhoon climate stage shift factor. Normalization calculations are performed based on the channel mean and channel standard deviation of the initial feature map to generate normalized initial values; The normalized initial value is adjusted based on the typhoon climate intensity scaling factor and the typhoon climate stage translation factor to generate the normalized feature; The conditional batch normalization unit includes a multilayer perceptron.

[0013] Furthermore, the process of generating multi-scale typhoon load abrupt change characteristics includes: The normalized features are passed through multi-scale convolutional sub-layers to generate multi-scale convolutional features; The normalized features are passed through a multi-scale convolutional gated mapping sub-layer to generate multi-scale gated weights; The initial multi-scale typhoon load mutation features are generated based on the element-wise product of the multi-scale convolutional features and multi-scale gating weights. The initial multi-scale typhoon load mutation features are passed through a convolutional compression layer to generate the multi-scale typhoon load mutation features. The multi-scale gated convolutional layer includes a multi-scale convolutional sub-layer, a multi-scale convolutional gated mapping sub-layer, and a convolutional compression layer.

[0014] Furthermore, the process of training the temporal feature capture model and generator using the wind speed load physical constraint loss function includes: Construct a physical constraint term for wind speed load based on the over-limit value of the predicted load exceeding the current wind speed load limit; Construct an adversarial loss term based on the authenticity of the load samples; The wind speed load physical constraint term and the countermeasure loss term are weighted and summed to generate the wind speed load physical constraint loss function.

[0015] Furthermore, the process of generating the load sample truth includes: The climate load temporal correlation features are copied along the time axis and then concatenated with the load sample sequence to generate comprehensive input features; The integrated input features are passed through a fully connected layer to generate the load sample authenticity. The dual-branch discriminator includes a fully connected layer.

[0016] Furthermore, the process of generating the predicted load includes: The integrated input features are passed through a transposed convolutional layer to generate the predicted load; The dual-branch discriminator includes a transposed convolutional layer.

[0017] Furthermore, the process of training the dual-branch discriminator based on the WGAN-GP loss function includes: A regression loss term is constructed based on the mean square error between the predicted load and the actual load; The WGAN-GP loss function is constructed by weighted summation of the regression loss term and the WGAN-GP loss term.

[0018] Compared with existing technologies, the beneficial effects of this invention are as follows: by integrating the temporal feature capture model of CNN and BiLSTM, and adopting the CGAN structure of convolutional generator and dual-branch discriminator, and combining wind speed load physical constraint loss and WGAN-GP loss function for joint training, this invention fully extracts the coupling relationship between typhoon climate characteristics and abnormal load sequences. This drives the CGAN generator to significantly expand the number of scenario samples required for risk assessment, making up for the prediction bias and risk omission caused by insufficient data in small samples and extreme scenarios. The dual-branch discriminator completes the load sample authenticity verification and predicted load output, realizing power load prediction adapted to the climate characteristics of typhoon areas and the situation of scarce typhoon samples, so as to generate a more accurate list of power market operation risks and improve the disaster resilience of the power market.

[0019] In particular, this invention achieves comprehensive and high-precision extraction of the correlation features between typhoon climate features and anomalous load sequences by integrating CNN and BiLSTM architectures and introducing a self-attention mechanism based on typhoon wind speed bias. The BiLSTM layer can deeply mine the bidirectional long-term temporal dependency between typhoon climate features and anomalous typhoon loads, accurately capturing the lag and nonlinear coupling patterns between the two. The self-attention mechanism adjusts the attention weights through an indicator function of the current typhoon wind speed and the typhoon effect threshold, which can specifically enhance the feature weights during the key impact period of the typhoon and weaken the interference of irrelevant time steps. The CNN convolutional layer focuses on extracting short-term local features of climate loads, making up for the shortcomings of BiLSTM in capturing local detailed features.

[0020] In particular, the generator of this invention is based on a convolutional neural network architecture, combining a fully connected reshaping layer, a conditional batch normalization unit, a multi-scale gated convolutional layer, and residual connections to achieve the generation of high-quality load sample sequences that fit the typhoon scenario. The conditional batch normalization unit has a built-in multilayer perceptron, which can accurately generate typhoon climate intensity scaling factor and typhoon climate stage translation factor by using the multilayer perceptron to make the normalization process closely fit the intensity changes and stage characteristics of the typhoon climate, ensuring that the generated load samples are always constrained by the typhoon scenario and completely avoiding the problem of samples deviating from actual meteorological conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the method for dynamically generating a power market operation risk list based on regional climate characteristics, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the time-series feature capture model of the dynamic generation method for the power market operation risk list based on regional climate characteristics, according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the generator of the method for dynamically generating a power market operation risk list based on regional climate characteristics, according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the dual-branch discriminator of the method for dynamically generating a power market operation risk list based on regional climate characteristics, according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0025] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] like Figures 1 to 4 As shown, this invention provides a method for dynamically generating a power market operation risk list based on regional climate characteristics. This method overcomes the problems in existing technologies where power load forecasting cannot adapt to the climate characteristics of typhoon areas and where typhoon samples are scarce. As a result, the power market operation risk list is not accurate enough and the power market has poor disaster resilience.

[0027] like Figure 1 As shown, this embodiment proposes a method for dynamically generating a power market operation risk list based on regional climate characteristics, including: The current typhoon climate feature sequence from the meteorological website and the typhoon abnormal load sequence from the power market trading platform are obtained. The current typhoon climate features and the typhoon abnormal load sequence are then used to generate climate load time correlation features through a time-series feature capture model based on CNN and BiLSTM architecture. The temporal correlation features of the climate load are used to generate a load sample sequence through a generator based on a convolutional neural network architecture; The load sample sequence and climate load temporal correlation features are processed through a dual-branch discriminator to generate load sample accuracy and predicted load, and a list of power market typhoon emergency operation risks is generated based on the predicted load. The generator and the dual-branch discriminator constitute a CGAN structure. The temporal feature capture model and the generator are trained using the wind speed load physical constraint loss function, and the dual-branch discriminator is trained based on the WGAN-GP loss function.

[0028] Specifically, Generative Adversarial Networks (GANs) and their variants mostly employ a single discriminator structure, which can only verify the authenticity of generated samples from a single dimension. They cannot simultaneously achieve the dual goals of sample authenticity assessment and accurate load prediction. Furthermore, traditional GANs suffer from problems such as training instability and pattern collapse, making it difficult to generate diverse and highly realistic load samples that fit the typhoon scenario, thus reducing the reliability of the risk list.

[0029] In response, this embodiment makes multiple improvements to the standard CGAN model structure to build a model that can accurately capture the nonlinear mapping relationship between meteorology and load under the condition of scarce historical typhoon data with very few samples. This provides a reliable, diverse and physically reasonable source-load prediction scenario for the risk list, thereby achieving quantitative early warning of power market risks.

[0030] Specifically, the process of generating a list of power market typhoon emergency operation risks based on the predicted load includes: The predicted load at multiple future time points is input into the electricity market clearing simulation program to obtain possible real-time electricity price sequences. If the proportion of times the real-time electricity price exceeds 1000 yuan / MWh in the total number of times is greater than 80%, then it is recorded in the electricity market typhoon emergency state operation risk list as: high risk of abnormal market price.

[0031] Based on the predicted load at multiple future time points, calculate the load fluctuation amplitude under multiple regional scenarios. and load ramp rate In the formula , Let represent the maximum and minimum predicted load values ​​for sample i in the same region at multiple future time points, respectively. This represents the predicted load for samples in the same region at future time point t+1 and future time point t, where future time point t is the time point closest to the current time. Based on whether the load fluctuation amplitude and load ramp-up rate exceed the system's reserve capacity threshold, and if the proportion of such scenarios exceeds a preset value, it is recorded in the power market typhoon emergency state operation risk list as a risk of power supply and demand imbalance.

[0032] The predicted load of the corresponding region is simulated by the Alternating Current Optimal Power Flow (AC-OPF) algorithm to simulate the dispatcher's optimal response after the disaster. Under the operating point solved by OPF, it is checked whether the power flow of the remaining lines exceeds its long-term or short-term emergency capacity limit, resulting in overload, and whether the voltage of each node exceeds the limit. The necessary load shedding amount to ensure that the power grid does not collapse is output, so as to record in the power market typhoon emergency state operation risk list: the system safety operation risk is high, and the expected load shedding amount of key transmission sections under the current forecast typhoon impact.

[0033] like Figure 2 As shown, the process of generating temporal correlation features of climate load further includes: The spliced ​​vector of the current typhoon climate feature sequence and the typhoon abnormal load sequence is passed through a preprocessing layer to generate a spliced ​​input vector; The concatenated input vector is passed through a BiLSTM layer to generate a long-term feature sequence of climate load. The long-term climate load feature sequence is used to generate enhanced time-step long-term climate load features through a typhoon wind speed bias self-attention mechanism. The concatenated input vector is passed through a convolutional layer to generate short-term climate load features; The enhanced time-step climate load long-term feature sequence and climate load short-term feature are concatenated along the channel dimension to generate the climate load time-related feature; The temporal feature capture model includes a preprocessing layer, a BiLSTM layer, a self-attention mechanism, and a convolutional layer.

[0034] like Figure 2 As shown, the process of generating enhanced time-step climate load long-term characteristics further includes: The time step elements of the long-term characteristic sequence of the climate load are respectively processed through a self-attention mechanism to generate initial time step attention weights; The indicator function based on the current typhoon wind speed and the typhoon effect threshold is multiplied by a bias coefficient to adjust the initial time step attention weights and generate prior bias-enhanced time step attention weights. The prior bias-enhanced time-step attention weights are used to weight and sum the time-step elements of the long-term climate load feature sequence to generate attention-enhanced features. The attention-enhanced features and time-step elements are concatenated into vectors to generate the enhanced time-step climate load long-term features.

[0035] Specifically, the process of generating temporal correlation features of climate load can be represented as follows: In the formula, X represents the long-term characteristic sequence of climate load, and X represents the current typhoon climate characteristic sequence. Indicates a BiLSTM layer. This indicates that the prior bias enhances the attention weights at each time step. express function, express function, , Let these represent the learnable convolutional weight matrix and the learnable convolutional bias term of the self-attention mechanism, respectively. This represents the bias coefficient, preferably 2.0 after cross-validation. Indicates the current typhoon wind speed With typhoon effect threshold The indicator function is set to 1 if the typhoon effect threshold is 10 m / s, and 0 otherwise. Indicating attention enhancement features, This indicates the long-term characteristics of enhanced time-step climate load. Indicates the short-term characteristics of climate load. Indicates the temporal correlation characteristics of climate load. This indicates that the long-term characteristic sequence of climate load and the short-term characteristic sequence of climate load at the enhanced time step are spliced ​​together along the channel dimension.

[0036] Specifically, the input current typhoon climate feature sequence includes the maximum wind speed and the average wind speed. The preprocessing layer includes the maximum and average wind speeds at multiple times, which are then normalized to their maximum and minimum values.

[0037] Understandably, different time steps contribute differently to load forecasting under the influence of typhoons. For example, the importance of the landfall time is much higher than that of the stable period. Therefore, this embodiment introduces an attention mechanism to dynamically allocate weights. By introducing prior experience, such as typhoon wind speeds of 10 m / s or higher causing a sudden drop in load, prior bias is introduced into the attention calculation to make the model pay more attention to high-risk periods with higher wind speeds.

[0038] It is understandable that the impact of typhoons on load includes both short-term drastic fluctuations, such as sudden wind speed changes, and long-term trend changes, such as the overall load decrease before and after the typhoon passes. Therefore, CNN and BiLSTM are used in parallel to extract multi-scale features. Experiments show that this multi-scale fusion method has superior performance in soft measurement, and it performs well in the soft measurement of the correlation features between typhoon climate and power grid load in this embodiment.

[0039] like Figure 3 As shown, the process of generating the load sample sequence further includes: The concatenated vector of the climate load time-related features and random noise is passed through a fully connected reshaping layer to generate an initial feature map; The initial feature map and the climate load time-related features are passed through a conditional batch normalization unit to generate normalized features; The normalized features are passed through a multi-scale gated convolutional layer to generate multi-scale typhoon load abrupt change features; After residual concatenation of the multi-scale typhoon load abrupt change features and the initial feature map, the load sample sequence is generated through a convolutional output layer. The generator includes a fully connected reshaping layer, a conditional batch normalization unit, a multi-scale gated convolutional layer, a residual connection layer, and a convolutional output layer.

[0040] Furthermore, the process of generating normalized features includes: The time-related features of the climate load are used through a multilayer perceptron to generate a typhoon climate intensity scaling factor and a typhoon climate stage shift factor. Normalization calculations are performed based on the channel mean and channel standard deviation of the initial feature map to generate normalized initial values; The normalized initial value is adjusted based on the typhoon climate intensity scaling factor and the typhoon climate stage translation factor to generate the normalized feature; The conditional batch normalization unit includes a multilayer perceptron.

[0041] Furthermore, the process of generating multi-scale typhoon load abrupt change characteristics includes: The normalized features are passed through multi-scale convolutional sub-layers to generate multi-scale convolutional features; The normalized features are passed through a multi-scale convolutional gated mapping sub-layer to generate multi-scale gated weights; The initial multi-scale typhoon load mutation features are generated based on the element-wise product of the multi-scale convolutional features and multi-scale gating weights. The initial multi-scale typhoon load mutation features are passed through a convolutional compression layer to generate the multi-scale typhoon load mutation features. The multi-scale gated convolutional layer includes a multi-scale convolutional sub-layer, a multi-scale convolutional gated mapping sub-layer, and a convolutional compression layer.

[0042] Specifically, the process of generating normalized features can be represented as: In the formula, Represents the initial feature map. This represents a feature reshaping operation, which reshapes a one-dimensional feature vector into a TxC feature map matrix, where T represents all historical time points and C represents all feature channels, corresponding to multiple input parameters and historical loads of typhoon climate. , These represent the learnable convolutional weight matrix and learnable convolutional bias term of the fully connected remodeling layer, respectively. This represents the concatenated vector of random noise z and the temporal correlation feature y of climate load. Represents normalized features. express Activation function , These represent the typhoon climate intensity scaling factor and the typhoon climate stage translation factor generated by the multilayer perceptron, respectively. The multilayer perceptron includes three layers with… Fully connected hidden layers with activation functions , These represent the channel mean and channel standard deviation of the initial feature map, respectively. This indicates element-wise multiplication. This represents the k-th scale feature of multi-scale convolution features. Represents a multi-scale convolutional sublayer. This represents the k-th scale weight of the multi-scale gating weights. This represents a multi-scale convolutional gated mapping sublayer. This represents the Sigmoid activation function. The initial multi-scale typhoon load abrupt change characteristics at the k-th scale are represented by k, which is preferably 3. That is, the three scale branches of the multi-scale convolutional sub-layer and the multi-scale convolutional gated mapping sub-layer adopt convolutional kernels of scales of 1×3, 1×5, and 1×7, respectively. This represents the initial multi-scale typhoon load abrupt change characteristics across all scales. This represents the first-scale convolutional features, the second-scale convolutional features, and the third-scale convolutional features. This indicates the characteristics of abrupt changes in typhoon load at multiple scales. This indicates a convolutional compression layer, using a 1×1 convolutional kernel to compress features to the target dimension. Indicates residual connectivity characteristics. This indicates that the multi-scale typhoon load abrupt change characteristics and the initial feature map are residually joined. Represents the load sample sequence. This represents the output layer of the convolution.

[0043] Therefore, in conditional batch normalization, the climate load time correlation features, which serve as the condition vector, are adjusted into scaling and translation factors through a multilayer perceptron in order to dynamically adjust the distribution of the normalized features, thereby enabling the generator to generate a load sequence that matches the current typhoon conditions.

[0044] Therefore, multi-scale gated convolution can simultaneously capture the dynamics of different time scales of load sequences under the influence of typhoons, and enhance key features through the gating mechanism.

[0045] Furthermore, the process of training the temporal feature capture model and generator using the wind speed load physical constraint loss function includes: Construct a physical constraint term for wind speed load based on the over-limit value of the predicted load exceeding the current wind speed load limit; Construct an adversarial loss term based on the authenticity of the load samples; The wind speed load physical constraint term and the countermeasure loss term are weighted and summed to generate the wind speed load physical constraint loss function.

[0046] Specifically, the wind speed load physical constraint loss function can be expressed as: In the formula, This represents the physical constraint loss function of wind speed load. This represents the adversarial loss term, i.e., the two-branch discriminator. The negative expectation of the generated sample scores, This represents the physical constraint term for wind speed load. Indicates the predicted load. Indicates the current wind speed The current wind speed-load ceiling is obtained by fitting a wind speed-load ceiling function to historical data. This represents the weighting coefficient, which is preferably 1.

[0047] like Figure 4 As shown, the process of generating the load sample truth value further includes: The climate load temporal correlation features are copied along the time axis and then concatenated with the load sample sequence to generate comprehensive input features; The integrated input features are passed through a fully connected layer to generate the load sample authenticity. The dual-branch discriminator includes a fully connected layer.

[0048] Specifically, the temporal correlation features of climate load are copied T times along the time axis to match the shape Tx1 of the load sample sequence, and then spliced ​​together. This fully connected layer includes two one-dimensional convolutional layers.

[0049] like Figure 4 As shown, the process of generating the predicted load further includes: The integrated input features are passed through a transposed convolutional layer to generate the predicted load; The dual-branch discriminator includes a transposed convolutional layer.

[0050] Specifically, at least two transposed convolutional layers can be progressively upsampled to the predicted time length, ultimately outputting a sequence of predicted loads.

[0051] Furthermore, the process of training the dual-branch discriminator based on the WGAN-GP loss function includes: A regression loss term is constructed based on the mean square error between the predicted load and the actual load; The WGAN-GP loss function is constructed by weighted summation of the regression loss term and the WGAN-GP loss term.

[0052] Specifically, the WGAN-GP loss function can be expressed as: In the formula, This represents the WGAN-GP loss function. This represents the WGAN-GP loss term. Represents the regression loss term. This indicates that the actual load x is distributed according to the actual data. After sampling, the expected value is calculated. Indicates the actual load. Indicates the predicted load. represents the weighting coefficient, preferably 2, so that the regression loss can provide stronger supervision when typhoon samples are scarce.

[0053] In this embodiment, by integrating the temporal feature capture model of CNN and BiLSTM, and adopting the CGAN structure of convolutional generator and dual-branch discriminator, and combining wind speed load physical constraint loss and WGAN-GP loss function for joint training, the coupling relationship between typhoon climate characteristics and abnormal load sequences is fully extracted. This drives the CGAN generator to significantly expand the number of scenario samples required for risk assessment, making up for the prediction bias and risk omission caused by insufficient data in small samples and extreme scenarios. The dual-branch discriminator completes the load sample authenticity verification and predicted load output, realizing power load prediction adapted to the climate characteristics of typhoon areas and the situation of scarce typhoon samples, so as to generate a more accurate list of power market operation risks and improve the disaster resilience of the power market. By integrating CNN and BiLSTM architectures and introducing a self-attention mechanism based on typhoon wind speed bias, a comprehensive and high-precision extraction of the correlation features between typhoon climate features and anomalous load sequences is achieved. The BiLSTM layer can deeply mine the bidirectional long-term temporal dependency between typhoon climate features and anomalous typhoon loads, accurately capturing the lag and nonlinear coupling patterns between the two. The self-attention mechanism adjusts the attention weights through an indicator function of the current typhoon wind speed and the typhoon effect threshold, which can specifically enhance the feature weights during the key impact periods of the typhoon and weaken the interference of irrelevant time steps. The CNN convolutional layer focuses on extracting short-term local features of the climate load, making up for the shortcomings of BiLSTM in capturing local detailed features. The generator is based on a convolutional neural network architecture, combining fully connected reshaping layers, conditional batch normalization units, multi-scale gated convolutional layers, and residual connections to achieve the generation of high-quality load sample sequences that fit the typhoon scenario. The conditional batch normalization unit has a built-in multilayer perceptron, which can first pass the temporal correlation features of the climate load through the multilayer perceptron to accurately generate typhoon climate intensity scaling factors and typhoon climate stage translation factors. This ensures that the normalization process closely fits the intensity changes and stage characteristics of the typhoon climate, ensuring that the generated load samples are always constrained by the typhoon scenario and completely avoiding the problem of samples deviating from actual meteorological conditions.

[0054] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0055] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamically generating a power market operation risk list tailored to regional climate characteristics, characterized in that, include: The current typhoon climate feature sequence from the meteorological website and the typhoon abnormal load sequence from the power market trading platform are obtained. The current typhoon climate features and the typhoon abnormal load sequence are then used to generate climate load time correlation features through a time-series feature capture model based on CNN and BiLSTM architecture. The temporal correlation features of the climate load are used to generate a load sample sequence through a generator based on a convolutional neural network architecture; The load sample sequence and climate load temporal correlation features are processed through a dual-branch discriminator to generate load sample accuracy and predicted load, and a list of power market typhoon emergency operation risks is generated based on the predicted load. The generator and the dual-branch discriminator constitute a CGAN structure. The temporal feature capture model and the generator are trained using the wind speed load physical constraint loss function, and the dual-branch discriminator is trained based on the WGAN-GP loss function.

2. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 1, characterized in that, The process of generating temporal correlation features of climate loads includes: The spliced ​​vector of the current typhoon climate feature sequence and the typhoon abnormal load sequence is passed through a preprocessing layer to generate a spliced ​​input vector; The concatenated input vector is passed through a BiLSTM layer to generate a long-term feature sequence of climate load. The long-term climate load feature sequence is used to generate enhanced time-step long-term climate load features through a typhoon wind speed bias self-attention mechanism. The concatenated input vector is passed through a convolutional layer to generate short-term climate load features; The enhanced time-step climate load long-term feature sequence and climate load short-term feature are concatenated along the channel dimension to generate the climate load time-related feature; The temporal feature capture model includes a preprocessing layer, a BiLSTM layer, a self-attention mechanism, and a convolutional layer.

3. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 2, characterized in that, The process of generating enhanced time-step climate load long-term characteristics includes: The time step elements of the long-term characteristic sequence of the climate load are respectively processed through a self-attention mechanism to generate initial time step attention weights; The indicator function based on the current typhoon wind speed and the typhoon effect threshold is multiplied by a bias coefficient to adjust the initial time step attention weights and generate prior bias-enhanced time step attention weights. The prior bias-enhanced time-step attention weights are used to weight and sum the time-step elements of the long-term climate load feature sequence to generate attention-enhanced features. The attention-enhanced features and time-step elements are concatenated into vectors to generate the enhanced time-step climate load long-term features.

4. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 1, characterized in that, The process of generating the payload sample sequence includes: The concatenated vector of the climate load time-related features and random noise is passed through a fully connected reshaping layer to generate an initial feature map; The initial feature map and the climate load time-related features are passed through a conditional batch normalization unit to generate normalized features; The normalized features are passed through a multi-scale gated convolutional layer to generate multi-scale typhoon load abrupt change features; After residual concatenation of the multi-scale typhoon load abrupt change features and the initial feature map, the load sample sequence is generated through a convolutional output layer. The generator includes a fully connected reshaping layer, a conditional batch normalization unit, a multi-scale gated convolutional layer, a residual connection layer, and a convolutional output layer.

5. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 4, characterized in that, The process of generating normalized features includes: The time-related features of the climate load are used through a multilayer perceptron to generate a typhoon climate intensity scaling factor and a typhoon climate stage shift factor. Normalization calculations are performed based on the channel mean and channel standard deviation of the initial feature map to generate normalized initial values; The normalized initial value is adjusted based on the typhoon climate intensity scaling factor and the typhoon climate stage translation factor to generate the normalized feature; The conditional batch normalization unit includes a multilayer perceptron.

6. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 4, characterized in that, The process of generating multi-scale typhoon load abrupt change characteristics includes: The normalized features are passed through multi-scale convolutional sub-layers to generate multi-scale convolutional features; The normalized features are passed through a multi-scale convolutional gated mapping sub-layer to generate multi-scale gated weights; The initial multi-scale typhoon load mutation features are generated based on the element-wise product of the multi-scale convolutional features and multi-scale gating weights. The initial multi-scale typhoon load mutation features are passed through a convolutional compression layer to generate the multi-scale typhoon load mutation features. The multi-scale gated convolutional layer includes a multi-scale convolutional sub-layer, a multi-scale convolutional gated mapping sub-layer, and a convolutional compression layer.

7. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 1, characterized in that, The process of training the time-series feature capture model and generator using the wind speed load physical constraint loss function includes: Construct a physical constraint term for wind speed load based on the over-limit value of the predicted load exceeding the current wind speed load limit; Construct an adversarial loss term based on the authenticity of the load samples; The wind speed load physical constraint term and the countermeasure loss term are weighted and summed to generate the wind speed load physical constraint loss function.

8. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 1, characterized in that, The process of generating load sample truth includes: The climate load temporal correlation features are copied along the time axis and then concatenated with the load sample sequence to generate comprehensive input features; The integrated input features are passed through a fully connected layer to generate the load sample authenticity. The dual-branch discriminator includes a fully connected layer.

9. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to claim 1, characterized in that, The process of generating predicted loads includes: The integrated input features are passed through a transposed convolutional layer to generate the predicted load; The dual-branch discriminator includes a transposed convolutional layer.

10. The method for dynamically generating a power market operation risk list based on regional climate characteristics according to any one of claims 1 to 9, characterized in that, The training process of the dual-branch discriminator based on the WGAN-GP loss function includes: A regression loss term is constructed based on the mean square error between the predicted load and the actual load; The WGAN-GP loss function is constructed by weighted summation of the regression loss term and the WGAN-GP loss term.

Images