A day-ahead electricity price prediction method and system fusing physical state identification and deep learning, a device, and a storage medium

Abstract

The application discloses a kind of day-ahead electricity price prediction method, system, equipment and storage medium of fusion physical state identification, relating to power market electricity price prediction technical field, including based on power market clearing mechanism, construct physical state variable, define the physical feasible region of price under specific operating state through multivariate kernel density estimation, provide state constraint for deep learning model;Physical perception loss function is constructed, and is embedded in the training process of deep learning model, guide the price sensitivity of deep learning model under different operating states is differentially modeled;Using the trained deep learning model, combined with the price distribution characteristics under physical state condition, the prediction sample deviating from the physical feasible region is constrained and corrected, and the day-ahead electricity price prediction result is obtained.The method disclosed in the application improves the physical consistency and explainability of the prediction result, strengthens the modeling ability of the model to extreme price intervals such as peak electricity price and negative electricity price, and improves the stability of the model.

Description

Technical Field

[0001] This invention relates to the field of electricity market price forecasting technology, specifically to a day-ahead electricity price forecasting method, system, device, and storage medium that integrates physical state identification and deep learning. Background Technology

[0002] With the deepening of the global energy structure transformation, the penetration rate of renewable energy sources such as wind power and photovoltaics in the power system continues to increase. While this trend has significantly reduced carbon emissions from the power system, it has also profoundly changed the physical supply and demand balance of the power system, making the formation mechanism of electricity market clearing prices more complex. Currently, day-ahead electricity prices in the current electricity spot market exhibit significant volatility, non-stationarity, and nonlinearity, with frequent extreme price events such as peak prices and negative prices. This not only significantly increases the decision-making risks for electricity market participants but also places higher demands on the safe and economical operation of the power system. Therefore, constructing a high-precision, robust day-ahead price prediction model that conforms to the physical operating mechanism of the power system has become crucial for research and engineering applications in the electricity market field.

[0003] Day-ahead electricity price forecasting is a typical multivariate, multi-step time series regression problem. Traditional statistical models and shallow machine learning methods, due to their limited feature extraction capabilities, struggle to capture the complex nonlinear patterns of electricity price series and are no longer sufficient to meet the electricity price forecasting needs under conditions of high-proportion renewable energy integration. In recent years, deep learning technology, with its powerful automatic feature mining capabilities, has become the mainstream technique for day-ahead electricity price forecasting. Related research mainly falls into two categories: one is feature representation enhancement methods, represented by Convolutional Neural Networks (CNNs) and Deep Belief Networks (DBNs), which can automatically extract statistical features, periodic features, and frequency domain information from electricity price time series; the other is time dependency modeling methods, represented by Recurrent Neural Networks (RNNs) and Long Short-Term Memory Networks (LSTMs), which can effectively characterize the temporal correlation and trend of electricity price series. Building on this foundation, generative models such as the Transformer architecture based on attention mechanisms and Generative Adversarial Networks (GANs) have also been applied to electricity price forecasting, further enhancing the model's ability to capture long-term time dependencies and electricity price distribution characteristics.

[0004] However, most existing deep learning-based day-ahead electricity price forecasting methods treat electricity price forecasting as a pure data fitting or time series regression problem, focusing on minimizing the global forecast error while neglecting the physical operating mechanism of the power system behind electricity market clearing. This belongs to a "physically unknowable" modeling paradigm, and such methods have significant defects in high-proportion renewable energy market environments, mainly in two aspects: First, the lack of physical constraints leads to a lack of physical consistency and interpretability in the forecast results. The electricity spot price is essentially a reflection of the system's supply and demand balance under the influence of unit ramp-up capabilities, start-up and shutdown characteristics. The results are the result of the combined effects of physical conditions such as power and capacity constraints and market clearing rules. Traditional deep learning models cannot explicitly perceive these hard boundaries, and are prone to outputting physically infeasible predictions in scenarios such as low load and high renewable energy output. Secondly, the mean regression tendency of traditional loss functions such as mean squared error (MSE) makes the model biased towards outputting smooth mean predictions, which makes it difficult to accurately characterize the nonlinear transition behavior of electricity prices in key ranges. Predictions for extreme price scenarios such as peak electricity prices and negative electricity prices are generally conservative, and there are problems of response lag and underestimation of magnitude. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is that existing deep learning-based day-ahead electricity price forecasting methods suffer from a lack of physical consistency in forecast results, conservative forecasting performance, and delayed response, as well as the problem of output results exceeding the physically feasible range.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a day-ahead electricity price forecasting method integrating physical state identification and deep learning, comprising: constructing physical state variables based on the electricity market clearing mechanism; defining the physical feasible domain of electricity prices under specific operating conditions through multivariate kernel density estimation, providing state constraints for the deep learning model; constructing a physical perception loss function and embedding it into the deep learning model training process to guide the deep learning model to perform differentiated modeling of electricity price sensitivity under different operating conditions; using the trained deep learning model, combined with the price distribution characteristics under physical state conditions, constraining and correcting the predicted samples that deviate from the physical feasible domain, and obtaining the day-ahead electricity price forecasting result.

[0008] As a preferred embodiment of the day-ahead electricity price forecasting method integrating physical state identification and deep learning described in this invention, the physical state variables include: system scarcity, renewable energy penetration rate, and net load ramp-up rate; the system scarcity is the ratio of net load to the available capacity of conventional thermal power units; the renewable energy penetration rate is the proportion of wind and solar renewable energy output to the total load; and the net load ramp-up rate is the ratio of the change in net load per unit time to the total capacity of thermal power units.

[0009] As a preferred embodiment of the day-ahead electricity price forecasting method integrating physical state identification and deep learning described in this invention, the physical feasible domain of electricity price under a specific operating state includes: constructing a physical state vector based on system scarcity, renewable energy penetration rate, and net load ramp-up rate; using a multivariate kernel density estimation method to perform nonparametric estimation of the joint probability density of the physical state and the clearing price; obtaining the conditional probability distribution of electricity price under a given state; and obtaining the physical feasible domain of electricity price by constructing a conditional cumulative distribution function and introducing a confidence level.

[0010] As a preferred embodiment of the day-ahead electricity price forecasting method integrating physical state identification and deep learning described in this invention, the joint probability density is expressed as: , in, For joint probability density, For the sample size, For those with bandwidth matrix Multivariate kernel functions, This is a four-dimensional composite observation vector, where the four dimensions are system scarcity, renewable energy penetration rate, net load ramp-up rate, and electricity price. For the first historical dataset One real sample.

[0011] As a preferred embodiment of the day-ahead electricity price forecasting method integrating physical state identification and deep learning described in this invention, the physical perception loss function includes a combination of a physical weighted fitting term and an out-of-domain boundary penalty term, specifically expressed as follows: , in, Let be the physical perception loss function. For physical sensitivity weights, for The actual clearing price at any given time for Electricity price forecast at time of day For balance coefficient, This is a penalty item for extraterritorial boundaries.

[0012] As a preferred embodiment of the day-ahead electricity price forecasting method integrating physical state identification and deep learning described in this invention, the method includes: differentiated modeling of electricity price sensitivity under different operating conditions, wherein the physical sensitivity weights are obtained based on the system scarcity and renewable energy penetration rate, with higher system scarcity resulting in greater weights and higher renewable energy penetration rates resulting in corresponding weight adjustments to reflect low-price characteristics; and an extra-domain boundary penalty term applies a secondary penalty to the predicted value that exceeds the physical feasible domain of the electricity price, thereby integrating marginal unit pricing theory into the loss function design and guiding the deep learning model to differentiate the sensitivity of the electricity price to changes under different operating conditions.

[0013] As a preferred embodiment of the day-ahead electricity price forecasting method integrating physical state identification and deep learning described in this invention, the constraint correction of the forecast samples deviating from the physical feasible region includes: statically truncating the original forecast value, and obtaining the correction value through a piecewise projection function in conjunction with the physical feasible region of the electricity price, specifically expressed as follows: , in, To determine the final correction value for the piecewise projection function, This is the lower limit of the electricity price. The original predicted value at time t, This is the upper limit of the electricity price.

[0014] Another objective of this invention is to provide a day-ahead electricity price forecasting system that integrates physical state identification and deep learning. This system can construct a physical perception loss function and embed it into the deep learning model training process to guide the deep learning model to perform differentiated modeling of electricity price sensitivity under different operating conditions. This solves the problem that current deep learning day-ahead electricity price forecasting lacks physical constraints, resulting in a lack of physical consistency in the forecasting results.

[0015] As a preferred embodiment of the day-ahead electricity price forecasting system integrating physical state identification and deep learning as described in this invention, the system includes: a physical state variable construction module, an electricity price feasible domain definition module, a loss function design module, a model training module, and a prediction result correction module. The physical state variable construction module is used to construct physical state variables that characterize the supply-demand tension and the dominance of new energy sources in the power system based on the power market clearing mechanism. The electricity price feasible domain definition module is used to define the physical feasible domain of electricity prices under specific operating conditions based on the physical state variables, providing a state constraint basis for the deep learning model. The loss function design module is used to design a physical perception loss function, incorporating marginal unit pricing theory into the loss function to guide the model to differentiate its sensitivity to price changes under different operating conditions. The model training module is used to embed the physical perception loss function into the training process of the deep learning model to train the deep learning model. The prediction result correction module is used to constrain and correct prediction samples that deviate from the physical feasible domain in the inference stage of the deep learning model, combining the price distribution characteristics under physical state conditions, to obtain the final day-ahead electricity price forecast result.

[0016] Another object of the present invention is to provide a day-ahead electricity price forecasting device that integrates physical state identification and deep learning, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the day-ahead electricity price forecasting method that integrates physical state identification and deep learning.

[0017] Another object of the present invention is to provide a day-ahead electricity price forecasting storage medium that integrates physical state identification and deep learning, wherein a computer program is stored thereon, and when the computer program is executed by a processor, the steps of the day-ahead electricity price forecasting method integrating physical state identification and deep learning are implemented.

[0018] The beneficial effects of this invention are: The day-ahead electricity price forecasting method provided by this invention integrates physical state identification and deep learning. Based on the dynamic definition of the physical feasible region of electricity prices using physical state variables, it transforms the marginal pricing mechanism of the electricity market into a physical hard constraint, effectively improving the physical consistency and interpretability of the forecast results. This avoids the model outputting physically infeasible electricity price forecasts. The physical perception loss function introduces physical sensitivity weights, differentially weighting the forecast errors for key scenarios such as system shortages and renewable energy dominance, mitigating the mean regression problem of traditional loss functions and strengthening the model's ability to characterize extreme price ranges such as peak and negative electricity prices. A two-layer error correction strategy of static truncation and dynamic correction is implemented to meet market requirements. Based on the absolute boundary of the field rules, the prediction results are corrected by combining the dynamic physical feasible region. Under the premise of minimal perturbation, the physical rationality of the prediction results is guaranteed, which improves the stability of the model under extreme conditions and out-of-distribution scenarios. It has good model independence and portability and can be applied to a variety of mainstream deep learning models such as Transformer, MLP, RNN, and CNN. It can improve the prediction accuracy and robustness of all of them. It is applicable to various electricity spot markets with a high proportion of new energy access and has broad engineering application value. This invention achieves better results in terms of prediction effect in extreme price scenarios, physical consistency of prediction results, and robustness and applicability of the method. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a day-ahead electricity price prediction method that integrates physical state identification and deep learning, as provided in Embodiment 1 of the present invention.

[0021] Figure 2 The image shows the prediction results of a day-ahead electricity price prediction method that integrates physical state identification and deep learning, as provided in Embodiment 2 of the present invention.

[0022] Figure 3 This is a schematic diagram of a day-ahead electricity price forecasting system that integrates physical state identification and deep learning, as provided in Embodiment 3 of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figure 1 As an embodiment of the present invention, a day-ahead electricity price prediction method integrating physical state identification and deep learning is provided, comprising: S1: Based on the electricity market clearing mechanism, a physical state variable 101 is constructed. The physical feasible domain 100 of electricity price under a specific operating state is defined by multivariate kernel density estimation, which provides state constraints for the deep learning model M.

[0025] Specifically, a physical state variable 101 is constructed to characterize the supply and demand tension of the power system and the dominance of new energy. The physical state variable 101 includes system shortage degree 1011, new energy penetration rate 1012 and net load ramp-up rate 1013, realizing the leap from simple data correlation mining to physical mechanism-driven modeling.

[0026] The physical state variables 101 include system scarcity 1011, renewable energy penetration rate 1012, and net load ramp-up rate 1013; the system scarcity 1011 is the ratio of net load to the available capacity of conventional thermal power units; the renewable energy penetration rate 1012 is the proportion of wind and solar renewable energy output to the total load; and the net load ramp-up rate 1013 is the ratio of the change in net load per unit time to the total capacity of thermal power units.

[0027] The system scarcity level of 1011 is specifically expressed as follows: , in, For system scarcity, for The load of time, for The constant output of new energy sources This represents the available capacity of conventional thermal power units.

[0028] The new energy penetration rate of 1012 is specifically expressed as follows: , in, This represents the penetration rate of new energy sources.

[0029] The net load ramp rate of 1013 is specifically expressed as follows: , in, Net load ramp-up rate, This represents the change in net load. The unit time interval.

[0030] By constructing the aforementioned physical state vectors, the deep learning model M is helped to move from simple data correlation mining to physical causal modeling. Specifically, it accurately characterizes the driving effect of marginal unit costs on electricity prices, effectively explaining the underlying physical causes of peak electricity prices, price collapses and negative value ranges, and scarcity pricing that deviates from marginal costs due to the shortage of flexible resources. Variables with clear economic and physical significance, combined with statistical learning methods, can define a reasonable electricity price range and eliminate long-tail noise interference from irrational electricity prices with extremely low probability.

[0031] Furthermore, defining the physical feasible region 100 of electricity price under a specific operating state includes constructing a physical state vector based on system scarcity 1011, renewable energy penetration rate 1012, and net load ramp-up rate 1013; using a multivariate kernel density estimation method to perform nonparametric estimation of the joint probability density of physical state and clearing price; obtaining the conditional probability distribution of electricity price under a given state; and obtaining the physical feasible region 100 of electricity price by constructing a conditional cumulative distribution function and introducing a confidence level.

[0032] Transforming the marginal pricing mechanism of the electricity market into an operational empirical physical hard constraint provides a state constraint foundation for the learning of the deep learning model M. First, a physical state space is constructed. Represented as: , in, for The physical state space at any given moment. For system scarcity, To increase the penetration rate of new energy sources, Net load ramp-up rate, This is a matrix transpose operation.

[0033] For physical state space The state vector in the equation represents the system at time t. The supply and demand balance and the marginal cost level of the generating units.

[0034] Secondly, the physical state is analyzed using the multivariate kernel density estimation method. and clearing electricity price The nonlinear coupling relationship between them is specifically achieved by aligning the historical physical state with the corresponding clearing price to reconstruct a four-dimensional composite observation vector. The kernel density estimation method using the multivariate Gaussian kernel function is selected for the joint probability density. Perform nonparametric estimation.

[0035] The joint probability density is expressed as: , in, For joint probability density, For the sample size, For those with bandwidth matrix Multivariate kernel functions, It is a four-dimensional composite observation vector. For the first historical dataset One real sample.

[0036] It should be noted that integrating the joint distribution along the electricity price dimension yields a given physical state. Conditional probability distribution of electricity price Specifically, it is expressed as: , in, physical state Conditional probability distribution of electricity price reduction physical state The lower joint probability density.

[0037] Finally, the conditional cumulative distribution function is constructed. Introducing confidence levels By truncating the long-tailed interference on both sides of the distribution, the cumulative probability reaches... and The electricity price at that time is used as the upper and lower bounds of the physically feasible region of the electricity price. The physical feasible region of the electricity price at time 100 is represented as: , in, for The physical feasible region of electricity price at any given time. This is the lower limit of the electricity price. This is the upper limit of the electricity price.

[0038] The conditional cumulative distribution function is specifically expressed as: , in, Let be the conditional cumulative distribution function. The effective integration range for clearing electricity prices is limited by the clearing price boundaries stipulated by the electricity markets of each province in practical applications.

[0039] The physical feasible region of electricity price, 100, is represented as follows: , in, For confidence level, This is the lower limit of the electricity price. This is the upper limit of the electricity price.

[0040] S2: Construct a physical perception loss function 201 and embed it into the training process of the deep learning model M to guide the deep learning model M to perform differentiated modeling of electricity price sensitivity under different operating conditions 200.

[0041] Specifically, a physical perception loss function 201 is designed, and the physical sensitivity weights are dynamically adjusted based on the system scarcity degree 1011 and the new energy penetration rate 1012, thereby differentiating the prediction error under key operating conditions.

[0042] The physics-based loss function 201 includes a combination of a physics-weighted fitting term and an out-of-domain boundary penalty term, specifically expressed as follows: , in, Let be the physical perception loss function. For physical sensitivity weights, for The actual clearing price at any given time for Electricity price forecast at time of day For balance coefficient, This is a penalty item for extraterritorial boundaries.

[0043] The physical sensitivity weight is expressed as follows: , in, For physical sensitivity weights, The weighting coefficients related to the system's scarcity level. The weighting coefficients related to the penetration rate of new energy sources It is an exponential function with the natural constant e as its base.

[0044] The extraterritorial boundary penalty term is represented as follows: , in, For extraterritorial boundary penalties, It is a function for maximizing the value.

[0045] Furthermore, the modeling of electricity price sensitivity under different operating conditions is differentiated 200. This includes obtaining the physical sensitivity weight based on the system scarcity 1011 and the renewable energy penetration rate 1012. The higher the system scarcity 1011, the greater the weight. The higher the renewable energy penetration rate 1012, the greater the weight is adjusted accordingly to reflect the low price characteristics. The extra-domain boundary penalty term applies a secondary penalty to the predicted value of the physical feasible domain 100 that exceeds the electricity price, thereby integrating the marginal unit pricing theory into the loss function design and guiding the deep learning model M to differentiate the sensitivity of electricity price changes under different operating conditions 200.

[0046] S3: Using the trained deep learning model M, combined with the price distribution characteristics under physical conditions, the predicted samples that deviate from the physical feasible region of 100 are constrained and corrected to obtain the day-ahead electricity price prediction result R.

[0047] Specifically, based on the designed physical perception loss function 201 embedded in the deep learning model M training process, the deep learning model M is trained using physical state variables 101, historical electricity prices, load forecasts, and new energy output forecasts as input features. The deep learning model M can be any one or more of Transformer, MLP, RNN, CNN and their improved variants, such as the improved variant of Transformer PatchTST, the improved variant of MLP TiDE, the improved variant of RNN xLSTM, and the improved variant of CNN TCN. All models use unified data input and prediction settings and are optimized within their respective recommended parameter configuration ranges to ensure the fairness of training.

[0048] During the inference phase of the deep learning model M, data is input into the trained deep learning model M to obtain the original electricity price prediction value. Then, the prediction samples that deviate from the physical feasible region of 100 are constrained and corrected by combining the price distribution characteristics under physical conditions to obtain the final day-ahead electricity price prediction result R.

[0049] The constraint correction for predicted samples deviating from the physical feasible region 100 includes statically truncating the original predicted values, combining the physical feasible region 100 of the electricity price, and obtaining the correction value through a piecewise projection function, specifically expressed as follows: , in, To determine the final correction value for the piecewise projection function, This is the lower limit of the electricity price. The original predicted value at time t, This is the upper limit of the electricity price.

[0050] Furthermore, the static truncation of the original forecast values ​​ensures that, regardless of the system's operating state, the electricity price forecast results should not exceed the absolute boundaries allowed by market rules. (Regarding time...) The original predicted values ​​are statically truncated, specifically as follows: , in, For time Static truncation of the original predicted values, It is a static truncation function. The original predicted value at time t, The absolute lower bound of electricity prices allowed by market rules. This is the absolute upper limit of the electricity price allowed by market rules.

[0051] Example 2, refer to Figure 2 As an embodiment of the present invention, a day-ahead electricity price prediction method integrating physical state identification and deep learning is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0052] First, data was collected using electricity spot market data from a certain province, with a time resolution of 15 minutes. Input variables included direct-dispatch load forecast, tie-line power forecast, total wind power output forecast, and total photovoltaic power output forecast. The data was divided into training, validation, and test sets according to time sequence. Baseline models were selected, including PatchTST (Transformer), TiDE (MLP), xLSTM (RNN), and TCN (CNN). A physically enhanced version incorporating the physical perception loss function 201 of this invention was also constructed. Evaluation metrics were set, using mean absolute error (MAE), root mean square error (RMSE), and weighted average absolute percentage error (WMAPE) to evaluate model performance. These three metrics comprehensively measure prediction accuracy from the perspectives of absolute error, sensitivity to outliers, and relative error.

[0053] Analysis of physical state variable 101 reveals that electricity prices do not fluctuate freely throughout the entire numerical space, but are significantly constrained by physical states such as system scarcity 1011 and renewable energy penetration rate 1012. When system scarcity 1011 increases, the physical feasible region 100 of electricity prices expands nonlinearly towards higher prices. When renewable energy penetration rate 1012 increases, the lower bound of the physical feasible region 100 of electricity prices shifts downward and even enters the negative range. The average electricity price exhibits a threshold effect with changes in system scarcity 1011 and a marginal decreasing characteristic with changes in renewable energy penetration rate 1012, verifying the rationality of the design of physical state variable 101.

[0054] Model comparison experiment results are as follows Figure 2As shown, after introducing the physical perception loss function 201 of this invention into each baseline model, the three evaluation indicators MAE, RMSE, and WMAPE all achieved varying degrees of performance improvement. Among them, the xLSTM model showed the most significant improvement, with MAE decreasing from 59.03 to 47.98, RMSE decreasing from 79.44 to 63.87, and WMAPE decreasing from 19.64% to 15.97%. The comparison of prediction results on typical high-fluctuation days shows that the method of this invention can respond to changes in the system's operating status more promptly, and its characterization of the timing and magnitude of price peaks is closer to the true clearing price, solving the problems of delayed response, underestimated amplitude, and over-smoothing in traditional models.

[0055] Ablation experiment: After sequentially introducing traditional exogenous variables, physical state variables 101, physical perception loss function 201, and the physical feasible region 100 constraint of electricity price, the model prediction error gradually decreased. When only historical electricity prices were used, the WMAPE was 35.66%, which decreased to 19.64% after introducing physical state variables 101, and to 16.36% after introducing physical perception loss function 201. Finally, after introducing the physical feasible region 100 constraint of electricity price, it decreased to 15.97%, and the abnormal predictions and physical violations were significantly reduced, verifying the effectiveness of each module of the present invention.

[0056] The experimental results above demonstrate that the present invention can significantly improve prediction accuracy and robustness in various deep learning models M, especially under extreme price scenarios, effectively verifying the effectiveness of explicitly embedding the physical mechanism of the power system into the deep learning model M to improve the reliability of electricity price prediction in a high proportion of new energy markets.

[0057] Example 3, referring to Figure 3 As an embodiment of the present invention, a day-ahead electricity price prediction system integrating physical state identification and deep learning is provided, including a physical state variable construction module, an electricity price feasible region definition module, a loss function design module, a model training module, and a prediction result correction module.

[0058] Among them, the physical state variable construction module is used to construct physical state variables 101 that characterize the supply and demand tension of the power system and the dominant characteristics of new energy, based on the power market clearing mechanism.

[0059] The electricity price feasible region definition module is used to define the physical feasible region 100 of electricity price under a specific operating state based on physical state variable 101, providing a state constraint basis for deep learning model M learning.

[0060] The loss function design module is used to design the physical perception loss function 201, which incorporates the marginal unit pricing theory into the loss function and guides the model to differentiate the sensitivity of the model to changes in electricity prices under different operating conditions 200.

[0061] The model training module is used to embed the physical perception loss function 201 into the training process of the deep learning model M, and to train the deep learning model M.

[0062] The prediction result correction module is used to correct the prediction samples that deviate from the physical feasible region of 100 by combining the price distribution characteristics under physical state conditions during the inference stage of the deep learning model M, and obtain the final day-ahead electricity price prediction result R.

[0063] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the day-ahead electricity price prediction method that integrates physical state identification and deep learning as proposed in the above embodiment.

[0064] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the day-ahead electricity price prediction method that integrates physical state identification and deep learning as proposed in the above embodiment.

[0065] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0067] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0068] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A day-ahead electricity price forecasting method integrating physical state identification and deep learning, characterized in that, include: Based on the power market clearing mechanism, physical state variables (101) are constructed, and the physical feasible domain (100) of electricity price under a specific operating state is defined by multivariate kernel density estimation, providing state constraints for the deep learning model (M); A physical perception loss function (201) is constructed and embedded into the training process of a deep learning model (M) to guide the deep learning model (M) to perform differentiated modeling of electricity price sensitivity under different operating conditions (200). Using the trained deep learning model (M), combined with the price distribution characteristics under physical state conditions, the predicted samples that deviate from the physical feasible region (100) are constrained and corrected to obtain the day-ahead electricity price prediction result (R).

2. The day-ahead electricity price forecasting method integrating physical state identification and deep learning as described in claim 1, characterized in that: The physical state variables (101) include, System scarcity (1011), new energy penetration rate (1012), and net load ramp-up rate (1013). The system shortage (1011) is the ratio of net load to the available capacity of conventional thermal power units; The new energy penetration rate (1012) is the proportion of wind and solar new energy output to the total load; The net load ramp rate (1013) is the ratio of the change in net load per unit time to the total capacity of the thermal power unit.

3. The day-ahead electricity price forecasting method integrating physical state identification and deep learning as described in claim 1 or 2, characterized in that: The physical feasible domain (100) defining the electricity price under a specific operating condition includes, Based on the system scarcity (1011), new energy penetration rate (1012), and net load ramp-up rate (1013), a physical state vector is constructed. The multivariate kernel density estimation method is used to perform nonparametric estimation of the joint probability density of the physical state and the clearing price, and the conditional probability distribution of the price under a given state is obtained. By constructing the conditional cumulative distribution function and introducing the confidence level, the physical feasible region (100) of the price is obtained.

4. The day-ahead electricity price forecasting method integrating physical state identification and deep learning as described in claim 3, characterized in that: The joint probability density is expressed as: ， in, For joint probability density, For the sample size, For those with bandwidth matrix Multivariate kernel functions, This is a four-dimensional composite observation vector, where the four dimensions are system scarcity, renewable energy penetration rate, net load ramp-up rate, and electricity price. For the first historical dataset One real sample.

5. The day-ahead electricity price forecasting method integrating physical state identification and deep learning as described in claim 1, 2, or 4, characterized in that: The physical perception loss function (201) includes, The combination of the physical weighted fitting term and the out-of-domain boundary penalty term is specifically expressed as: ， in, Let be the physical perception loss function. For physical sensitivity weights, for The actual clearing price at any given time for Electricity price forecast at time of day For balance coefficient, This is a penalty item for extraterritorial boundaries.

6. The day-ahead electricity price forecasting method integrating physical state identification and deep learning as described in claim 5, characterized in that: The differentiated modeling of electricity price sensitivity under different operating conditions (200) includes, The physical sensitivity weight is obtained based on the system scarcity (1011) and the new energy penetration rate (1012). The higher the system scarcity (1011), the greater the weight. The higher the new energy penetration rate (1012), the greater the weight is adjusted accordingly to reflect the low price characteristics. The extra-domain boundary penalty term applies a secondary penalty to the predicted value of the physical feasible domain (100) that exceeds the electricity price, thereby integrating the marginal unit pricing theory into the loss function design and guiding the deep learning model (M) to differentiate the sensitivity of the electricity price change under different operating conditions (200).

7. The day-ahead electricity price forecasting method integrating physical state identification and deep learning as described in claim 1, 2, 4, or 6, characterized in that: The constraint correction for the predicted samples that deviate from the physical feasible region (100) includes, The original predicted value is statically truncated, and the corrected value is obtained through a piecewise projection function, combined with the physical feasible region (100) of the electricity price. Specifically, it is expressed as follows: ， in, To determine the final correction value for the piecewise projection function, This is the lower limit of the electricity price. The original predicted value at time t, This is the upper limit of the electricity price.

8. A day-ahead electricity price forecasting system integrating physical state identification and deep learning, employing the day-ahead electricity price forecasting method integrating physical state identification and deep learning as described in any one of claims 1 to 7, characterized in that: It includes a physical state variable construction module, an electricity price feasible region definition module, a loss function design module, a model training module, and a prediction result correction module; The physical state variable construction module is used to construct physical state variables (101) that characterize the supply and demand tension of the power system and the dominant characteristics of new energy, based on the power market clearing mechanism. The electricity price feasible domain definition module is used to define the physical feasible domain (100) of electricity price under a specific operating state based on physical state variables (101), providing a state constraint basis for deep learning model (M) learning; The loss function design module is used to design the physical perception loss function (201), which incorporates the marginal unit pricing theory into the loss function and guides the model to differentiate the sensitivity of the model to changes in electricity price under different operating conditions (200). The model training module is used to embed the physical perception loss function (201) into the training process of the deep learning model (M) to train the deep learning model (M); The prediction result correction module is used to perform constraint correction on the prediction samples that deviate from the physical feasible region (100) in the inference stage of the deep learning model (M) by combining the price distribution characteristics under physical state conditions, so as to obtain the final day-ahead electricity price prediction result (R).

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the day-ahead electricity price forecasting method that integrates physical state identification and deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the day-ahead electricity price prediction method that integrates physical state identification and deep learning as described in any one of claims 1 to 7.

Images