Multi-region adaptive GRU power market electricity price rolling prediction method and system

Abstract

The present application relates to the technical field of power market prediction, in particular to a multi-region adaptive GRU power market electricity price rolling prediction method and system, first, a multi-dimensional standardized input sequence integrating time weight modulation is constructed, which is input into the corresponding three-layer GRU model of the region to generate the next time point electricity price standardized prediction value, and the actual electricity price is obtained through reverse standardization and time consistency and supply-demand consistency correction, then the supply-demand implicit feature component is extracted from the actual electricity price, the input sequence is rolling updated after reorganizing and standardizing the features, and the high-frequency electricity price prediction sequence of the preset time span is generated in a cycle; the present application solves the problems of insufficient region adaptation capability, lack of reverse feature optimization mechanism and multi-step prediction error accumulation in the prior art, significantly improves the accuracy and stability of multi-region high-frequency electricity price prediction, and can effectively support power system operation optimization, market transaction decision and risk control.

Description

Technical Field

[0001] This invention relates to the field of electricity market forecasting technology, specifically to a rolling forecasting method and system for electricity market prices using a multi-regional adaptive GRU. Background Technology

[0002] With the advancement of electricity market liberalization and the implementation of high-frequency trading mechanisms, coupled with the integration of a high proportion of renewable energy, electricity prices exhibit significant temporal volatility and regional differences. The dynamic changes in supply and demand directly impact the electricity price formation mechanism. Therefore, high-frequency electricity price forecasting for short timescales has gradually become a crucial foundational technology for power system operation optimization, market trading decisions, and risk control. This technology involves the integration of power system analysis, time series modeling, and intelligent forecasting algorithms. It requires models that can not only handle multidimensional heterogeneous market characteristics but also possess a deep understanding of the supply-demand coupling relationship, thereby achieving accurate predictions of electricity price trends.

[0003] Traditional technologies mainly employ multivariate joint modeling or graph structure learning methods to predict electricity prices. For example, Chinese patent CN119674967B provides a method and system for joint prediction of wind power, load and electricity prices in multiple regions. This method constructs a multi-scale graph convolutional network to uniformly model wind power output, load demand and electricity prices in multiple regions, and realizes information interaction between multiple target variables through shared feature layers, thereby improving the overall prediction accuracy.

[0004] These methods focus on utilizing spatial correlation and multivariate coupling to jointly learn data from multiple regions through a unified model structure, which enhances the model's ability to express the characteristics of complex systems to some extent. However, they still have significant shortcomings in high-frequency electricity price rolling forecasting scenarios. The unified modeling approach based on multiple regions is insufficient to fully characterize the differences in supply and demand structures and price formation mechanisms in the electricity markets of different regions, limiting the model's adaptability at the regional level. At the same time, existing methods focus on forward feature learning and lack a mechanism to extract supply and demand information from the forecast results, resulting in the inability to dynamically optimize the input features. Furthermore, the lack of an effective rolling closed-loop update structure in the multi-step forecasting process easily leads to error accumulation problems, making it difficult to support high-frequency stable forecasting over long periods. Therefore, in view of the above situation, there is an urgent need to develop a multi-regional adaptive GRU method and system for rolling electricity market price forecasting to overcome the shortcomings in current practical applications. Summary of the Invention

[0005] The purpose of this invention is to provide a rolling forecasting method and system for electricity market prices using a multi-region adaptive GRU, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The multi-region adaptive GRU method for rolling electricity market price forecasting includes the following steps: S1: Construct a standardized input sequence incorporating time-weighted modulation by utilizing the latest collected multi-dimensional real market characteristics; S2: Input the standardized input sequence into an independently trained three-layer GRU model corresponding to the electricity market region to which the prediction task belongs, and infer to generate the standardized predicted value of electricity price for the next time point; S3: Perform destandardization on the standardized electricity price forecast, and perform time consistency correction and supply-demand consistency correction to obtain the actual electricity price forecast; S4: Extract implicit feature components related to supply and demand from the actual electricity price forecast; S5: Recombine the hidden feature components with multi-dimensional market features into a new feature vector, and then update the input sequence after standardization; Repeat steps S2 to S5 to generate a complete sequence of high-frequency electricity price rolling forecasts covering a preset time span.

[0007] As a further aspect of the present invention: step S1 specifically includes: Acquire multi-dimensional market characteristic data for multiple consecutive time points prior to the current moment. The multi-dimensional market characteristics include at least aggregate demand, initial supply, clearing supply, dispatchable generation, renewable energy ratio, and standby service prices with different response speeds. Perform Z-score normalization transformation on each feature dimension separately; A time-weighted modulation mechanism is introduced to apply time decay weights to the standardized features, so that data closer to the current time has a greater impact on the model. The weighted features from multiple consecutive time points are stacked in chronological order to construct a standardized input sequence matrix.

[0008] As a further aspect of the present invention: In step S2, the three-layer GRU model, from bottom to top, consists of a first GRU unit layer, a first dropout regularization layer, a second GRU unit layer, a second dropout regularization layer, a third GRU unit layer, a third dropout regularization layer, and a fully connected output layer. During inference, the first and second layer GRUs output the complete hidden state sequence for each time step, while the third layer GRU only outputs the hidden state vector for the last time step. After being mapped by the fully connected output layer, the standardized predicted value of the electricity price for the next time point is generated. A regional feature modulation factor is introduced in the output mapping stage to correct the amplitude of the prediction results in order to adapt to the market volatility characteristics of different regions.

[0009] As a further aspect of this invention: a state consistency modulation mechanism is introduced in the second-layer GRU to smooth the hidden states of continuous time steps. The calculation formula is as follows: ; in, This indicates that the second-layer GRU is in the... The hidden state vector at each time step, This indicates the output of the first layer at the corresponding time step. Represents the state fusion coefficient. This indicates the hidden state of the second layer at the previous time step.

[0010] As a further aspect of the present invention: step S3 specifically includes: Call the standardizer corresponding to the target electricity price, perform the inverse transformation operation of Z-score standardization, and map the standardized predicted value of the electricity price back to the absolute value space of the original electricity price; A time consistency correction mechanism is introduced to dynamically adjust the current forecast result based on the actual electricity price value of the previous time step; A supply-demand consistency constraint is introduced, and the forecast results are structurally corrected based on the ratio of the total demand forecast to the total supply capacity at the next time point.

[0011] As a further aspect of the present invention: step S4 specifically includes: Extract the total demand forecast for the next time point from the historical electricity market forecast database; Based on the inherent correlation between actual electricity price forecasts and supply-side characteristics, we reverse-decompose and extract implicit characteristic components that reflect the degree of tension in the balance between electricity supply and demand. The hidden feature components are subjected to continuous enhancement processing to suppress fluctuation noise in the single-step inversion process.

[0012] As a further aspect of the present invention: the formula for extracting the hidden feature components is as follows: ; in, This represents the implicit characteristic components of supply and demand. This represents the actual electricity price forecast output from step three. This represents the average electricity price parameter. This represents the demand forecast. Indicates the effective capacity of the supply side. This represents the supply-demand coupling sensitivity coefficient.

[0013] As a further aspect of the present invention: step S5 specifically includes: The implicit feature components are combined and recombined with the multi-dimensional market features at the same prediction time point to form a complete feature vector. The recombined feature vectors are standardized, and a distribution alignment mechanism is introduced to ensure that the statistical properties of the new input features are consistent with those of the current input sequence. The standardized feature vectors are concatenated to the end of the input sequence, while the data from the earliest time step is removed to maintain a fixed-length time window structure.

[0014] A multi-region adaptive GRU-based rolling electricity market price forecasting system includes: The data preprocessing module is used to construct a standardized input sequence incorporating time-weighted modulation by utilizing the latest collected multi-dimensional real market characteristics; The regional GRU inference module is used to input the standardized input sequence into an independently trained three-layer GRU model corresponding to the electricity market region to which the prediction task belongs, and to infer and generate a standardized prediction value of the electricity price at the next time point. The destandardization correction module is used to perform destandardization on the electricity price standardization forecast value, and perform time consistency correction and supply and demand consistency correction to obtain the actual electricity price forecast value. The implicit feature decomposition module is used to decompose and extract implicit feature components related to supply and demand from the actual electricity price forecast; The rolling update module is used to recombine the hidden feature components with multi-dimensional market features into a new feature vector, and then roll-up update the input sequence after standardization. The control module controls the operation of the cyclic execution region GRU inference module, the denormalization correction module, the hidden feature decomposition module, and the rolling update module to generate a complete high-frequency electricity price rolling forecast sequence covering a preset time span.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By constructing a multi-dimensional standardized input sequence incorporating time-weighted modulation, the multi-dimensional market features achieve scale consistency and temporal enhancement before entering the model, significantly reducing the impact of different dimensions and noise interference on model inference, thereby improving the stability and short-term dynamic response capability of high-frequency electricity price forecasting. A three-layer GRU model structure with independent training in multiple regions is adopted, and regional feature modulation factors are combined to adapt to the differences in market structure in different regions. This enables the model to adaptively model for the differences in electricity market structure, effectively enhancing the generalization ability and accuracy of the prediction model in cross-regional scenarios. By performing destandardization on the electricity price forecast and overlaying time consistency and supply-demand consistency corrections, the forecasted electricity price can simultaneously satisfy statistical distribution characteristics and market economic laws, avoiding abnormal fluctuations in the forecast results, thereby improving the reliability of electricity price forecasts in actual trading and dispatch applications. We introduce a supply and demand implicit feature decomposition method based on electricity price inversion. This method extracts the supply and demand implicit feature components from the electricity price forecast and performs smoothing to enhance the information integrity of the input features, enabling the model to more accurately characterize the impact of supply and demand tension on electricity prices. By using a rolling feature recombination and fixed-length time window update mechanism, the prediction results are dynamically fed back to subsequent inputs, forming a closed-loop optimization structure. This effectively suppresses error accumulation during continuous multi-step prediction, thereby achieving stable output of high-frequency electricity price prediction over a long period of time. Attached Figure Description

[0016] Figure 1 This is a flowchart of the rolling forecasting method for electricity market prices using a multi-region adaptive GRU in an embodiment of the present invention.

[0017] Figure 2 This is a diagram of the multi-dimensional feature input sequence structure in an embodiment of the present invention.

[0018] Figure 3 This is a diagram of the three-layer GRU network hierarchy in an embodiment of the present invention.

[0019] Figure 4 This is a multi-layer correction logic diagram for electricity price forecasts in an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of the time window rolling iteration update in an embodiment of the present invention.

[0021] 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.

[0022] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0023] Please see Figures 1-5 The present invention provides a method and system for rolling electricity market price forecasting based on multi-region adaptive GRU, which solves the technical problems in the prior art such as insufficient multi-region unified modeling adaptation capability, lack of reverse feature optimization mechanism, and accumulation of multi-step prediction error. It can achieve high-precision and high-stability rolling electricity market price forecasting, and meet the core needs of power system operation optimization, market transaction decision-making and risk control.

[0024] like Figure 1 As shown in the embodiment of the present invention, the multi-region adaptive GRU method for rolling electricity market price forecasting generally includes the following steps: Step S1: Construct a standardized input sequence for the first prediction using the latest collected multi-dimensional real market features; Step S2: Input the standardized input sequence into the three-layer GRU model corresponding to the region, and infer to generate the standardized predicted value of electricity price for the next time point; Step S3: Perform destandardization on the standardized electricity price forecast to restore it to the actual electricity price value. Step S4: Retrieve the demand forecast value for the next time point from the historical data of the same period, and decompose the implicit characteristic components related to supply and demand from the actual electricity value; Step S5: Reorganize the hidden feature components and the demand prediction value into a new multidimensional feature vector, and after standardization, roll it to the end of the input sequence to drive the prediction at the next time point; Repeat steps S2 to S5 to generate a complete sequence of high-frequency electricity price rolling forecasts covering a preset time span.

[0025] The specific implementation methods for each step are described in detail below: Step S1: Construct the normalized input sequence for the first prediction Before initiating the rolling forecasting process, complete multi-dimensional market characteristic data for five consecutive time points prior to the current moment are obtained from the electricity market data source. These multi-dimensional market characteristics include at least nine dimensions of feature information: aggregate demand, initial supply, clearing supply, dispatchable generation, renewable energy proportion, and standby service prices with four different response speeds. In this embodiment, the time granularity for each time point is 30 minutes, which is consistent with the settlement cycle of high-frequency electricity market transactions, meeting the time accuracy requirements for high-frequency electricity price forecasting.

[0026] For each feature dimension, a pre-persisted normalizer is invoked. Using the mean and standard deviation of that feature dimension recorded in the normalizer on the training set, Z-score normalization is performed dimension-by-dimensionally. Z-score normalization is a commonly used data normalization method in this field. It transforms the original data into standard normally distributed data with a mean of 0 and a standard deviation of 1, eliminating the influence of different units on model training and inference. For the j-th dimension market feature at the i-th time point, the normalization calculation process is as follows: ; in, Indicates the first The first time point Original market characteristic values, Indicates the first The global mean of the dimensional features is obtained by statistical analysis during the model training phase. This represents the corresponding standard deviation. Represents the standardized eigenvalues; The purpose of this process is to eliminate the dimensional differences between different features, so that all input variables are on a uniform scale, thereby avoiding gradient bias problems caused by inconsistent numerical ranges in the GRU gating structure, while ensuring that the model's sensitivity to each feature remains consistent.

[0027] After completing dimension-by-dimensional standardization, to enhance the dynamic expressive power of the time series, a time weight modulation mechanism is introduced when constructing the five-step time window input sequence. This mechanism allows data closer to the current time to have a greater impact on the model. The calculation formula is as follows: ; in, This represents the standardized eigenvalues ​​after time decay weighting. This represents the time step from the current time, and its value ranges from zero to four. This represents the time decay coefficient, which is obtained through optimization on the validation set during the training phase. This represents the original standardized result; The purpose of this formula is to strengthen the dominant role of near-time information in prediction through an exponential decay mechanism, while suppressing noise accumulation in far-time information, thereby improving the response sensitivity of the GRU model in capturing short-term electricity price fluctuations.

[0028] After time-weighting, the nine-dimensional features from five consecutive time points are stacked in chronological order to form a matrix, constructing a unified input tensor structure, the structure of which is as follows: ; in, The final standardized input sequence matrix is ​​represented by rows representing the time dimension and columns representing the feature dimension. The overall structure is five by nine. This structure not only maintains strict consistency of the time order, but also encodes time decay information through weighted feature values. This enables the input sequence to have dynamic trend expression ability and cross-dimensional collaborative feature distribution consistency before entering the three-layer GRU model. This significantly reduces the difficulty of model adaptation under different regional data distribution differences and provides a stable and high-information-density input foundation for regional model selection and sequence inference in subsequent steps.

[0029] Step S2: Generate standardized electricity price forecasts using a three-layer GRU model for regional inference. Based on the electricity market region identifier to which the current forecasting task belongs, load the corresponding three-layer gated recurrent unit neural network model from five pre-trained regional independent model libraries. For example... Figure 3As shown, the network structure of this model, from bottom to top, consists of a first GRU unit layer, a first dropout regularization layer, a second GRU unit layer, a second dropout regularization layer, a third GRU unit layer, a third dropout regularization layer, and a fully connected output layer. Each GRU unit layer has 30 units, and the retention probability of each dropout regularization layer is 0.8. Dropout regularization is a commonly used technique in this field to prevent overfitting in neural networks. By randomly discarding the output of some neurons during training, it enhances the model's generalization ability.

[0030] The standardized input sequence constructed in step S1 is fed into the model. The first and second layer GRUs output the complete hidden state sequence of each time step for the next layer to continue processing. The third layer GRU only outputs the hidden state vector of the last time step. After the hidden state vector is mapped by the fully connected output layer, the standardized predicted value of the electricity price at the first 30-minute interval in the future is generated.

[0031] In the first layer of GRU, the hidden state at each time step is determined by the current input and the hidden state at the previous time step, and its calculation formula is as follows: ; in, This represents the time step index, with values ​​ranging from zero to four. This indicates the first step after time weighting. Nine-dimensional feature vectors at each time step This indicates that the first-layer GRU is in the... The hidden state vector at each time step, This indicates updating the gate vector. This indicates resetting the gate vector. and These represent the input weight matrix and the state weight matrix, respectively. This represents element-wise multiplication. The purpose of this computation process is to adaptively fuse short-term fluctuations and historical dependencies in the input sequence through a gating mechanism, so that important features are preserved while redundant information is suppressed, thereby constructing a hidden representation sequence with temporal memory capabilities.

[0032] In the second-layer GRU, the input consists of the complete hidden state sequence output from the first layer. To further enhance the stability of information across time steps, a state consistency modulation mechanism is introduced in this layer to smooth the hidden states in consecutive time steps. The calculation formula is as follows: ; in, This indicates that the second-layer GRU is in the... The hidden state vector at each time step, This indicates the output of the first layer at the corresponding time step. This represents the state fusion coefficient, which is determined during the region model training phase. This indicates the hidden state of the second layer at the previous time step; The purpose of this formula is to reduce the high-frequency disturbances in the input sequence caused by market mutations through joint constraints across layers and time steps, so that the hidden state remains smooth and continuous in the time dimension, thereby improving the stability and noise resistance of the model in high-frequency electricity price prediction.

[0033] In the third-layer GRU, only the hidden state at the final time step is retained as the global temporal representation vector. To enhance the expressive power of market structure differences in different regions, a regional feature modulation factor is introduced in the output mapping stage, and its calculation formula is as follows: ; in, This represents the standardized forecast of electricity prices for the next time step. This represents the hidden state vector of the third-layer GRU at the last time step. This represents the weight vector of the fully connected output layer. Indicates the current electricity market area identification. The corresponding regional modulation coefficients, The second norm of a vector; The purpose of this formula is to introduce a regional-level amplitude adjustment term in addition to the basic linear mapping, so that the model can adjust the amplitude of the prediction results according to the market fluctuation characteristics of different regions, thereby enhancing the generalization consistency and prediction accuracy of the model in multi-regional scenarios, and finally obtaining the electricity price prediction results in the standardized space.

[0034] Step S3: De-standardization operation to restore the value of electricity under the actual electricity price scale The system invokes the target standardizer corresponding to the target electricity price dimension. This standardizer persistently stores the mean and standard deviation parameters of the target electricity price variable from the model training phase. The standardized predicted value of the electricity price is multiplied by the target electricity price standard deviation parameter and then added to the target electricity price mean parameter. An inverse Z-score standardization transformation is then performed, mapping the predicted value output by the model in the standardized space back to the absolute value space of the original electricity price. The core de-standardization process is as follows: ; in, This represents the actual electricity price forecast after adjustment. This represents the standardized electricity price forecast output from step two. This represents the mean parameter of the target electricity price obtained during the model training phase. This represents the corresponding standard deviation parameter; The purpose of this formula is to map the prediction results in the standardized space back to the original electricity price scale, so that it is restored to a value with practical settlement significance, while maintaining consistency with the distribution of training data, thereby avoiding the impact of scale shift on subsequent business decisions.

[0035] After completing the basic inverse standardization, to ensure the smoothness and stability of the predicted electricity price in a continuous time series, a time consistency correction mechanism is introduced to dynamically adjust the current prediction results. The calculation formula is as follows: ; in, This represents the electricity price forecast after time consistency correction. and These represent the actual electricity price values ​​for the first two time steps, respectively. Indicates the time-trend modulation coefficient; This parameter is determined by the regional model during the training phase. The purpose of this formula is to compensate the current forecast value by introducing historical price change trends, so that the forecast results maintain continuous change characteristics in the time dimension, avoid abrupt changes caused by single-step prediction of the model, and thus enhance the overall stability of the rolling forecast sequence.

[0036] Furthermore, to ensure that the predicted electricity price is inherently consistent with the supply and demand structure, a supply and demand consistency constraint term is introduced in the output stage to perform structural correction on the prediction results. The calculation formula is as follows: ; in, This represents the final output of the actual electricity price forecast. This indicates the electricity price after time-consistency correction. This represents the total demand forecast for the next time step. This represents the total supply capacity at the same time step. Indicates the supply and demand sensitivity coefficient; This parameter reflects the strength of the electricity price response to the degree of supply and demand imbalance. The purpose of this formula is to structurally adjust the predicted electricity price by explicitly introducing the supply-demand ratio, so that it not only conforms to the statistical distribution, but also to the basic economic law that supply and demand determine prices in the electricity market, thereby significantly improving the credibility and interpretability of the prediction results in practical application scenarios.

[0037] Through the above three-stage processing, a final electricity price forecast with physical meaning and economic consistency is obtained. This electricity price is expressed in the actual settlement currency unit of the Australian National Electricity Market and can be directly used in downstream business scenarios such as trading decisions, risk control, and energy dispatch.

[0038] Step S4: Decompose and extract the implicit feature components related to supply and demand. Based on the time identifier corresponding to the current forecast time, the total demand forecast value consistent with the future forecast time is retrieved and extracted from the historical electricity market forecast database for the same period that has been acquired and stored in advance. This demand forecast value reflects the expected level of electricity consumption at the next time point.

[0039] Simultaneously, based on the actual electricity price forecast obtained in step S3, and according to the inherent economic relationship between electricity prices and supply and demand fundamentals in the electricity market, the forecast is jointly calculated with currently known supply-side characteristics such as dispatchable generation and the proportion of renewable energy. This process extracts implicit feature components reflecting the tightness of the electricity supply-demand balance at that point in time. These implicit feature components encode supply and demand relationship information contained within the electricity price forecast that is not directly expressed by other explicit features.

[0040] First, to achieve a reverse mapping from electricity prices to supply and demand structure, an effective capacity representation on the supply side is introduced, which integrates the calculation of dispatchable generation and the proportion of renewable energy. The calculation formula is as follows: ; in, Indicates the effective supply capacity on the supply side. This indicates the dispatchable power generation at the next point in time. Indicates the proportion of renewable energy. This indicates the conversion factor for renewable energy output; This parameter reflects the equivalent supply contribution of renewable energy under unstable conditions. The purpose of this formula is to uniformly characterize the contribution of different types of power sources to actual supply capacity, so that supply-side characteristics can participate in subsequent inversion calculations on a single scale, thereby providing a consistent basic dimension for the decomposition of implicit characteristics.

[0041] After constructing the key variables on both the supply and demand sides, based on the inherent coupling relationship between electricity prices and supply-demand balance, a structural inversion is performed on the actual electricity price forecast to extract the implicit expression of the supply-demand tension. The calculation formula is as follows: ; in, This represents the implicit characteristic components of supply and demand. This represents the actual electricity price forecast output from step three. This represents the average electricity price parameter. This represents the demand forecast. Indicates the effective capacity of the supply side. Indicates the sensitivity coefficient between supply and demand; This parameter is determined during the regional model training phase. The purpose of this formula is to extract the remaining information as a latent feature component by logarithmically normalizing the electricity price and removing the part explained by the explicit supply-demand ratio. This allows the latent feature component to specifically characterize the supply-demand tension in the electricity price that is not covered by directly observed variables, thereby enhancing the expressive completeness of subsequent model inputs.

[0042] Furthermore, to improve the stability and transitivity of the latent feature over time, a continuity enhancement process is applied to the latent feature component, the calculation formula of which is as follows: ; in, This represents the smoothed hidden feature components. This represents the latent feature calculated at the current time. This indicates the implicit features of the previous moment. This parameter represents the feature update weight and controls the fusion ratio of old and new information. The purpose of this formula is to suppress fluctuation noise in the single-step inversion process, so that the latent features can maintain continuous evolution characteristics in the rolling prediction process, thereby improving their stability and usability in subsequent input reconstruction.

[0043] Step S5: Feature Recombination and Rolling Update of Input Sequence The implicit feature components obtained in step S4 are concatenated and recombined with the demand forecast values ​​at the same prediction point in time along the feature dimensions. Simultaneously, the remaining seven dimensions of market feature values—initial supply, clearing supply, dispatchable generation, renewable energy ratio, and prices of four types of reserve services—are supplemented from historical data for that point in time. These are arranged in the same nine-dimensional feature order as in step S1, forming a complete nine-dimensional feature vector for the future prediction point. The calculation formula is as follows: ; in, This represents the recombined nine-dimensional feature vector. Indicates the initial supply. Indicates the clearing of supply. Indicates the dispatchable power generation. Indicates the proportion of renewable energy. to This indicates the prices for three different types of backup services with varying response speeds. This represents the implicit characteristic components of supply and demand; The purpose of this formula is to encode explicit supply and demand variables and implicit supply and demand information in a unified manner, so that the input features include both observable market variables and structural information obtained from electricity price inversion, thereby improving the explanatory power of the input features for future electricity price changes.

[0044] The nine-dimensional feature vector is standardized dimension-by-dimensionally using the same dimensional normalizer as in step S1, and a consistent alignment mechanism is introduced to ensure the stability of the statistical distribution during the rolling process. The calculation formula is as follows: ; in, Represents the standardized first 3D eigenvalues This represents the corresponding original feature value. and They represent the first The mean and standard deviation of the dimensional features. This represents the standardized value of the dimension corresponding to the end of the current input sequence. This represents the average value of that dimension in the current sequence. Indicates the distribution alignment coefficient; The purpose of this formula is to fine-tune the distribution of new input features while performing standardization, so that it is consistent with the overall statistical characteristics of the current sequence, thereby avoiding a decrease in model inference stability due to input distribution drift during rolling prediction.

[0045] Subsequently, the standardized feature vectors are concatenated to the end of the input sequence, and the data from the earliest time step is removed simultaneously to maintain a fixed time window structure. This rolling update process is achieved through sequence transformation, and its calculation formula is as follows: ; in, This represents the updated, normalized input sequence. This represents the subsequence from the second to the fifth time step in the original input sequence. This represents the newly generated standardized feature vector, and the symbol indicates the concatenation operation of the time dimension. The purpose of this formula is to achieve rolling updates of information within a fixed-length time window, so that the model always performs inference based on the data from the latest five time steps, while preserving the continuity of historical information.

[0046] By continuously executing the above feature recombination and sequence update process, and forming a closed-loop iteration with steps S2 to S4, the model can gradually advance towards future time. In this embodiment, steps S2 to S5 are executed a total of 30 times, ultimately generating a complete high-frequency electricity price rolling forecast sequence covering the next 15 hours with a time granularity of 30 minutes.

[0047] This invention also provides a multi-region adaptive GRU-based rolling forecasting system for electricity market prices, used to implement the above method. The system includes: The data preprocessing module is used to perform the operation of step S1, and to construct a standardized input sequence for the first prediction using the latest collected multi-dimensional real market characteristics. The regional GRU inference module is used to perform the operation in step S2, inputting the standardized input sequence into the three-layer GRU model corresponding to the region, and inferring to generate the standardized predicted value of electricity price for the next time point. The destandardization correction module is used to perform the operation in step S3, which performs a destandardization operation on the electricity price standardization prediction value to restore it to the electricity price value under the actual electricity price scale. The implicit feature decomposition module is used to perform the operation of step S4, retrieve the demand forecast value for the next time point from the historical data of the same period, and decompose the implicit feature components related to supply and demand from the actual electricity value. The rolling update module is used to perform the operation of step S5, which reorganizes the hidden feature components and the demand prediction value into a new multidimensional feature vector, and after standardization, it is rolled and spliced ​​to the end of the input sequence to drive the prediction of the next time point. The control module controls the operation of the cyclic execution region GRU inference module, the denormalization correction module, the hidden feature decomposition module, and the rolling update module to generate a complete high-frequency electricity price rolling forecast sequence covering a preset time span.

[0048] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A rolling forecasting method for electricity market prices using a multi-region adaptive GRU, characterized in that, Includes the following steps: S1: Construct a standardized input sequence incorporating time-weighted modulation by utilizing the latest collected multi-dimensional real market characteristics; S2: Input the standardized input sequence into an independently trained three-layer GRU model corresponding to the electricity market region to which the prediction task belongs, and infer to generate the standardized predicted value of electricity price for the next time point; S3: Perform destandardization on the standardized electricity price forecast, and perform time consistency correction and supply-demand consistency correction to obtain the actual electricity price forecast; S4: Extract implicit feature components related to supply and demand from the actual electricity price forecast; S5: Recombine the hidden feature components with multi-dimensional market features into a new feature vector, and then update the input sequence after standardization; Repeat steps S2 to S5 to generate a complete sequence of high-frequency electricity price rolling forecasts covering a preset time span.

2. The rolling forecasting method for electricity market prices using a multi-region adaptive GRU according to claim 1, characterized in that, Step S1 specifically includes: Acquire multi-dimensional market characteristic data for multiple consecutive time points prior to the current moment. The multi-dimensional market characteristics include at least aggregate demand, initial supply, clearing supply, dispatchable generation, renewable energy ratio, and standby service prices with different response speeds. Perform Z-score normalization transformation on each feature dimension separately; A time-weighted modulation mechanism is introduced to apply time decay weights to the standardized features, so that data closer to the current time has a greater impact on the model. The weighted features from multiple consecutive time points are stacked in chronological order to construct a standardized input sequence matrix.

3. The rolling forecasting method for electricity market prices using a multi-region adaptive GRU according to claim 1, characterized in that, In step S2, the three-layer GRU model consists of, from bottom to top, a first GRU unit layer, a first dropout regularization layer, a second GRU unit layer, a second dropout regularization layer, a third GRU unit layer, a third dropout regularization layer, and a fully connected output layer. During inference, the first and second layer GRUs output the complete hidden state sequence for each time step, while the third layer GRU only outputs the hidden state vector for the last time step. After being mapped by the fully connected output layer, the standardized predicted value of the electricity price for the next time point is generated. A regional feature modulation factor is introduced in the output mapping stage to correct the amplitude of the prediction results in order to adapt to the market volatility characteristics of different regions.

4. The rolling forecasting method for electricity market prices using a multi-region adaptive GRU according to claim 3, characterized in that, A state consistency modulation mechanism is introduced in the second-layer GRU to smooth the hidden states at continuous time steps. The calculation formula is as follows: ； in, This indicates that the second-layer GRU is in the... The hidden state vector at each time step, This indicates the output of the first layer at the corresponding time step. Represents the state fusion coefficient. This indicates the hidden state of the second layer at the previous time step.

5. The rolling forecasting method for electricity market prices using a multi-region adaptive GRU according to claim 1, characterized in that, Step S3 specifically includes: Call the standardizer corresponding to the target electricity price, perform the inverse transformation operation of Z-score standardization, and map the standardized predicted value of the electricity price back to the absolute value space of the original electricity price; A time consistency correction mechanism is introduced to dynamically adjust the current forecast result based on the actual electricity price value of the previous time step; A supply-demand consistency constraint is introduced, and the forecast results are structurally corrected based on the ratio of the total demand forecast to the total supply capacity at the next time point.

6. The rolling forecasting method for electricity market prices using a multi-region adaptive GRU according to claim 1, characterized in that, Step S4 specifically includes: Extract the total demand forecast for the next time point from the historical electricity market forecast database; Based on the inherent correlation between actual electricity price forecasts and supply-side characteristics, we reverse-decompose and extract implicit characteristic components that reflect the degree of tension in the balance between electricity supply and demand. The hidden feature components are subjected to continuous enhancement processing to suppress fluctuation noise in the single-step inversion process.

7. The rolling forecasting method for electricity market prices using a multi-region adaptive GRU according to claim 6, characterized in that, The formula for extracting the hidden feature components is as follows: ； in, This represents the implicit characteristic components of supply and demand. This represents the actual electricity price forecast output from step three. This represents the average electricity price parameter. This represents the demand forecast. Indicates the effective capacity of the supply side. This represents the supply-demand coupling sensitivity coefficient.

8. The rolling forecasting method for electricity market prices using a multi-region adaptive GRU according to claim 1, characterized in that, Step S5 specifically includes: The implicit feature components are combined and recombined with the multi-dimensional market features at the same prediction time point to form a complete feature vector. The recombined feature vectors are standardized, and a distribution alignment mechanism is introduced to ensure that the statistical properties of the new input features are consistent with those of the current input sequence. The standardized feature vectors are concatenated to the end of the input sequence, while the data from the earliest time step is removed to maintain a fixed-length time window structure.

9. A rolling forecasting system for electricity market prices based on a multi-regional adaptive GRU, characterized in that, include: The data preprocessing module is used to construct a standardized input sequence incorporating time-weighted modulation by utilizing the latest collected multi-dimensional real market characteristics; The regional GRU inference module is used to input the standardized input sequence into an independently trained three-layer GRU model corresponding to the electricity market region to which the prediction task belongs, and to infer and generate a standardized prediction value of the electricity price at the next time point. The destandardization correction module is used to perform destandardization on the electricity price standardization forecast value, and perform time consistency correction and supply and demand consistency correction to obtain the actual electricity price forecast value. The implicit feature decomposition module is used to decompose and extract implicit feature components related to supply and demand from the actual electricity price forecast; The rolling update module is used to recombine the hidden feature components with multi-dimensional market features into a new feature vector, and then roll-up update the input sequence after standardization. The control module controls the operation of the cyclic execution region GRU inference module, the denormalization correction module, the hidden feature decomposition module, and the rolling update module to generate a complete high-frequency electricity price rolling forecast sequence covering a preset time span.

Images