Intelligent Trading Methods and Devices for the Electricity Market

Abstract

This disclosure relates to a smart trading method and apparatus for the electricity market. The smart trading method includes: acquiring historical trading data, power grid physical state data, wind farm data, and external environmental data from the electricity market; determining a market state feature vector based on the acquired data; and generating trading actions for a predetermined time period based on the market state feature vector through a policy-evaluation network incorporating an attention mechanism. The trading actions include submitting electricity demand and submitting electricity price. The reward function applied during the training of the policy-evaluation network includes a core revenue term, a deviation penalty term, a risk penalty term, and a constraint reward term. By adopting this disclosure, decision-making efficiency and accuracy can be improved, market adaptability can be enhanced, and returns can be optimized. It better adapts to the characteristics of large daily rolling market fluctuations and complex market information, improves computational efficiency, and adopts a distributed training architecture to support parallel processing of large-scale historical data.

Description

Technical Field

[0001] This disclosure relates to the field of power trading technology, and more specifically, to a smart trading method and apparatus for the power market. Background Technology

[0002] Currently, trading strategies are primarily formulated based on traditional electricity price trend forecasting. The limitations of traditional cost-based pricing methods in electricity trading lie in their static and inward-looking nature. They rarely consider the strategies of other market participants, and when numerous power generation companies adopt similar "cost-plus" pricing models, it can lead to homogeneous competition. This can result in the inability to obtain reasonable high-price profits during periods of strong demand, or the loss of market opportunities due to a lack of pricing flexibility during periods of intense competition. Essentially, it reflects a production-oriented mindset rather than a market-oriented one.

[0003] Market-based electricity price forecasting optimization methods represent a significant advancement over cost analysis, paying closer attention to market dynamics. However, their core problem lies in passive adaptation. These methods typically assume their trading volume is insufficient to influence market prices, positioning them as "price takers." However, in real strategic bidding, the actions of large power generation companies themselves alter the market supply and demand dynamics, thus affecting clearing prices. Furthermore, market-based electricity price forecasting optimization methods are highly dependent on forecast accuracy. With the substantial increase in the proportion of highly volatile renewable energy sources such as wind and solar power, the difficulty and uncertainty of electricity price forecasting have increased dramatically; a single forecasting error can lead to significant deviation assessment costs or opportunity losses.

[0004] Game theory equilibrium analysis is theoretically perfect, attempting to simulate the strategic interactions of all market participants to find the optimal solution. However, its key bottleneck lies in information. Real-world electricity markets suffer from severe information asymmetry; competitors' costs, next-day power generation plans, and actual bidding strategies are all private information. The strict assumption of complete information is difficult to satisfy in real-world markets. Furthermore, the computation of solving multi-player game equilibria is extremely complex, making it difficult to use in day-ahead and intraday markets that require rapid responses.

[0005] The high proportion of renewable energy integration has fundamentally altered the operating rules of the electricity market. Its volatility, intermittency, and low marginal cost characteristics have led to significantly increased market price volatility and uncertainty. The negative electricity prices observed in many regions during specific periods in recent years are a new situation that traditional theories struggle to explain. This means that traditional methods relying on historical patterns and stability assumptions are inadequate for dealing with these new market scenarios. Trading strategies based on electricity price trend forecasts are only guided by short-term predictions and cannot comprehensively consider factors such as trading risk and long-term return expectations to output the trading strategy with the best risk-reward ratio. Summary of the Invention

[0006] To address the aforementioned issues, this disclosure proposes an intelligent trading method and apparatus, a computing system, and a computer-readable storage medium for the electricity market.

[0007] According to one aspect of this disclosure, a smart trading method for the electricity market is provided. The smart trading method for the electricity market includes: acquiring historical trading data of the electricity market, power grid physical state data, wind farm data, and external environment data, and determining a market state feature vector based on the acquired data; and generating trading actions for a predetermined time period based on the market state feature vector through a policy-evaluation network incorporating an attention mechanism, wherein the trading actions include submitting electricity demand and submitting electricity price, and the reward function applied during the training process of the policy-evaluation network includes a core benefit term, a deviation penalty term, a risk penalty term, and a constraint reward term.

[0008] Optionally, the step of determining the market state feature vector based on the acquired data includes: selecting a predetermined number of historical transaction data from the historical transaction data of the electricity market based on Euclidean distance, with the timestamps closest to the current transaction date, to form a sample set; determining key features based on the sample set, the key features including at least a supply and demand tension index, price fluctuation statistics, and time series features; and determining the market state feature vector based on the key features and real-time observed electricity market transaction data, power grid physical state data, and wind farm data.

[0009] Optionally, the output of the strategy-evaluation network includes the mean and standard deviation of the declared electricity price and declared electricity volume, as well as the probability distribution of the transaction action determined based on the mean and standard deviation.

[0010] Optionally, the core revenue item is determined based on the declared electricity price, the cleared electricity price, and the cleared electricity volume; the deviation penalty item is determined based on the difference between the declared electricity volume and the cleared electricity volume; and the risk penalty item is determined based on the value at risk at a preset confidence level and the current expected rate of return, calculated from historical revenue data of similar days.

[0011] Optionally, the training process of the strategy-evaluation network includes: mapping the market state feature vector into a query vector, a key vector, and a value vector through an attention layer, and determining a weighted feature vector based on the query vector, key vector, and value vector using a normalized exponential function; inputting the weighted feature vector into the strategy network and the evaluation network included in the strategy-evaluation network, respectively; generating the trading action through the strategy network; determining the value assessment of the trading action through the evaluation network; and updating the parameters of the strategy network and the evaluation network through backpropagation of network gradients based on the reward function and the value assessment.

[0012] Optionally, mean squared error is used as the loss function during the backpropagation of the network gradient.

[0013] Optionally, both the policy network and the evaluation network include three fully connected hidden layers.

[0014] According to another aspect of this disclosure, a smart trading device for an electricity market is provided, the smart trading device comprising: a market state feature vector determination unit configured to acquire historical trading data of the electricity market, power grid physical state data, wind farm data, and external environment data, and determine a market state feature vector based on the acquired data; and a trading action generation unit configured to generate trading actions for a predetermined time period based on the market state feature vector through a strategy-evaluation network incorporating an attention mechanism, wherein the trading actions include submitting electricity volume and submitting electricity price, and the reward function applied during the training process of the strategy-evaluation network includes a core benefit term, a deviation penalty term, a risk penalty term, and a constraint reward term.

[0015] Optionally, the market state feature vector determination unit is further configured to: select a predetermined number of historical transaction data from the historical transaction data of the electricity market based on Euclidean distance, with the timestamp closest to the current transaction date, to form a sample set; determine key features based on the sample set, the key features including at least a supply and demand tension index, price fluctuation statistics, and time series features; and determine the market state feature vector based on the key features and real-time observed electricity market transaction data, power grid physical state data, and wind farm data.

[0016] Optionally, the output of the strategy-evaluation network includes the mean and standard deviation of the declared electricity price and declared electricity volume, as well as the probability distribution of the transaction action determined based on the mean and standard deviation.

[0017] Optionally, the core revenue item is determined based on the declared electricity price, the cleared electricity price, and the cleared electricity volume; the deviation penalty item is determined based on the difference between the declared electricity volume and the cleared electricity volume; and the risk penalty item is determined based on the value at risk at a preset confidence level and the current expected rate of return, calculated from historical revenue data of similar days.

[0018] Optionally, the training process of the strategy-evaluation network includes: mapping the market state feature vector into a query vector, a key vector, and a value vector through an attention layer, and determining a weighted feature vector based on the query vector, key vector, and value vector using a normalized exponential function; inputting the weighted feature vector into the strategy network and the evaluation network included in the strategy-evaluation network, respectively; generating the trading action through the strategy network; determining the value assessment of the trading action through the evaluation network; and updating the parameters of the strategy network and the evaluation network through backpropagation of network gradients based on the reward function and the value assessment.

[0019] Optionally, mean squared error is used as the loss function during the backpropagation of the network gradient.

[0020] Optionally, both the policy network and the evaluation network include three fully connected hidden layers.

[0021] According to another aspect of this disclosure, a computing system is provided that includes at least one computing device and at least one storage device for storing instructions, wherein the instructions, when executed by the at least one computing device, cause the at least one computing device to perform the smart trading method for the electricity market as described above.

[0022] According to another aspect of this disclosure, a computer-readable storage medium is provided that stores instructions, wherein when the instructions are executed by at least one computing device, they cause the at least one computing device to perform the smart trading method for the electricity market as described above.

[0023] By adopting this disclosure, decision-making efficiency can be improved, enabling 24 / 7 monitoring of multi-source market data (e.g., prices, load, renewable energy output, etc.) and instantaneous completion of the entire process from analysis to decision-making, seizing short-term arbitrage opportunities that are difficult to capture manually; decision-making accuracy can be improved by approximating complex market state mappings through deep neural networks, overcoming the dependence of traditional methods on model simplification, and achieving near-optimal strategies in environments with incomplete information; market adaptability can be enhanced by having online learning capabilities, dynamically adjusting pricing strategies according to market price fluctuations and policy changes, adapting to the market uncertainty in scenarios with a high proportion of renewable energy; returns can be optimized by not pursuing absolute accuracy in single price predictions, but rather by comprehensively considering risks and long-term effects to maximize the optimal trading strategy for long-term cumulative returns, better adapting to the characteristics of large daily rolling market fluctuations and complex market information; and computational efficiency can be improved by adopting a distributed training architecture, supporting parallel processing of large-scale historical data, and solving the curse of dimensionality problem in traditional game theory methods. Attached Figure Description

[0024] The above and / or other objects and advantages of this disclosure will become clearer from the following description of embodiments in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating an exemplary embodiment of a smart trading method for an electricity market according to the present disclosure; Figure 2 This is a diagram illustrating a strategy-evaluation network structure according to an embodiment of the present disclosure; Figure 3 This is a block diagram illustrating an intelligent trading device for an electricity market according to an exemplary embodiment of the present disclosure; Figure 4 This is a block diagram illustrating a computing system including at least one computing device and at least one storage device of storage instructions according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0025] The following description, in conjunction with the accompanying drawings, provides specific embodiments to aid the reader in gaining a comprehensive understanding of the methods, apparatus, and / or systems described herein. However, upon understanding this disclosure, various changes, modifications, and equivalents of the methods, apparatus, and / or systems described herein will become apparent. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but may be altered as will become clear upon understanding this disclosure, except for operations that must occur in a specific order. Furthermore, for clarity and conciseness, descriptions of features known in the art may be omitted.

[0026] This disclosure proposes a method for constructing an intelligent agent for electricity spot market trading based on policy-evaluation networks and deep reinforcement learning. The trading agent can adopt a layered architecture, including, for example, an environment perception module and an intelligent decision-making module. The environment perception module integrates multi-source data such as grid dispatch data (node ​​prices, line power flow), weather forecasts, load predictions, and unit operating status; it uses an attention mechanism to dynamically weight multi-dimensional features and extracts key influencing factors of the market state, forming the foundation for the agent's perception of the market state. It uses a time-series database to store timestamped market data. Intelligent Decision Module: Based on a strategy-evaluation framework incorporating an attention mechanism; an attention mechanism module is embedded before each network. Through the attention mechanism, the model can dynamically learn and assign different weights to different features (for example, the weight of meteorological factors will automatically increase when the weather changes suddenly; the weight of topology and power flow data will increase significantly when the line is blocked), enabling the agent to flexibly focus on the core contradictions of the current market like an experienced trader; the strategy network takes market state features as input and outputs pricing actions; the evaluation network evaluates the value of actions and evaluates the strategy score through optimization algorithms. Its goal is to learn a function that maps the market state to the optimal pricing strategy; at the same time, a trading strategy risk term is embedded to calculate the Value at Risk (VaR) in real time, which constrains the agent's actions. The VaR measures how confident (e.g., 90% confidence) the loss will not exceed a certain value in a specific future time under normal market conditions.

[0027] Figure 1 This is a flowchart illustrating an exemplary embodiment of a smart trading method for an electricity market according to the present disclosure. Figure 2 This is a diagram illustrating a strategy-evaluation network structure according to an embodiment of the present disclosure.

[0028] like Figure 1 As shown, in step S101, historical transaction data of the electricity market, power grid physical status data, wind farm data and external environment data are acquired, and market status feature vectors are determined based on the acquired data.

[0029] In the example, a predetermined number of historical transaction data with timestamps closest to the current transaction date are selected from the historical transaction data of the electricity market based on Euclidean distance to form a sample set; based on the sample set, key features are determined, including at least the supply and demand tension index, price fluctuation statistics, and time series features; based on the key features and real-time observed electricity market transaction data, power grid physical state data, and wind farm data, a market state feature vector is determined.

[0030] For example, it integrates historical market data (e.g., day-ahead / real-time electricity prices, cleared electricity volume, etc.), grid physical status (e.g., line power flow, congestion, etc.), substation data (e.g., unit output, ramp rate, energy consumption, etc.), and external environmental data (e.g., weather forecasts accurate to the grid, holiday information, etc.). Based on the data source dimension, Euclidean distance is used to filter the daily data closest to the current transaction date from historical data as the data filter set. The Euclidean distance calculation formula is shown below.

[0031] Suppose there are two vectors , The distance is ,but .

[0032] For example, the following key features can be constructed: Supply and demand tension index: (System forecast load - New energy forecast output) / Total system supply capacity; Price fluctuation characteristics: Calculate the rolling mean, standard deviation, and Pearson median skewness of historical electricity prices; Temporal characteristics: marking time period, day of the week, whether it is a holiday, etc.

[0033] For example, state This involves observing the electricity trading market and needs to include all the information required for decision-making. The state vector can be designed as follows:

[0034] in, and Let be the day-ahead electricity price and the real-time electricity price at time t, respectively. This is the load forecast at time t. and This is the prediction of new energy power output at time t. This is the power output prediction for the current station at time t. It is the blocking state vector of each adjacent node of the power grid at time t (the value is 1 when it is in a blocking state and 0 when it is in a non-blocking state, forming a zero-one vector). It is the transaction status at time t-1 (e.g., it may include information such as open position, declared electricity price, declared electricity volume, and profit amount).

[0035] In step S102, based on the market state feature vector, a strategy (Actor)-critic network incorporating an attention mechanism is used to generate trading actions for a predetermined time period. The trading actions include submitting electricity demand and submitting electricity price. The reward function applied during the training of the strategy-critic network includes a core revenue term, a deviation penalty term, a risk penalty term, and a constraint reward term.

[0036] For example, actions For decision-making in trading models, the action space is typically continuous for continuous bidding problems.

[0037]

[0038] in, It is the declared electricity price (yuan / MWh). This refers to the declared electricity consumption (MWh). To ensure the feasibility of the action, constraints must be imposed: , ,in, P min and P max These are the lower and upper limits for the declared electricity price, respectively. Q max This is the upper limit for the declared electricity volume (based on the unit capacity).

[0039] In the example, the core revenue item is determined based on the declared electricity price, the cleared electricity price, and the cleared volume; the deviation penalty item is determined based on the difference between the declared volume and the cleared volume; and the risk penalty item is determined based on the value at risk at a preset confidence level and the current expected rate of return, calculated from historical revenue data of similar days.

[0040] For example, reward function In other words, the Target serves as the basis for the agent's decision-making and can be designed to balance immediate gains, long-term risks, and physical constraints. Its core structure is as follows:

[0041]

[0042] set up The design of the reward function is essentially based on the expectation of the current policy. Infinitely close to the expected reward of the long-term strategy That is, the optimization objective function of the algorithm. This is the discount factor, which can be set to 0.95. , To evaluate the network's output value after training, i.e., the reward at the current time step and the reward at the next time step.

[0043] Core benefits: That is, the actual profit difference in the current period ((declared electricity price - clearing electricity price) × clearing volume).

[0044] Deviation penalty items: The penalty is the discrepancy between the declared electricity volume and the actual cleared electricity volume.

[0045] , Adjust the severity of penalties to encourage accurate pricing and avoid market scrutiny for inaccurate predictions.

[0046] Risk penalty items: Introducing Value at Risk (VaR) measurement at a given confidence level (Here, it's set to 90%) the maximum possible loss; Calculate the return rate for each similar day, where P represents the transaction price and Q represents the transaction volume. It is a variable The relevant risk-return calculation function is used here, and then the returns are arranged in ascending order to form a return series, and the 1- That is, the return value corresponding to the 10th percentile (the worst 10% return); This aims to discourage agents from adopting overly aggressive strategies and guide them to pursue risk-adjusted returns.

[0047] Similar days are historical dates that are highly similar to the target prediction date in key features (e.g., day of the week, weather conditions, holidays, lunar calendar solar terms, etc.). Common similarity calculation methods include weighted Euclidean distance, grey relational analysis, and cosine similarity. A typical process for determining similar days may include: constructing feature vectors for historical days and prediction days; calculating the similarity between all candidate historical days and prediction days; sorting by similarity and selecting the top N most similar days (e.g., N=3 or 7); and using the historical data of these similar days as prediction input.

[0048] Constraints on reward items: Positive rewards are given when the agent's strategy satisfies physical constraints such as unit ramp rate and output limits. Furthermore, the index function is 1 when the constraints are met, and 0 otherwise.

[0049] In the example, the training process of the strategy-evaluation network includes: mapping market state feature vectors to query vectors, key vectors, and value vectors through an attention layer, and determining weighted feature vectors based on the query vectors, key vectors, and value vectors using a normalized exponential function; inputting the weighted feature vectors into the strategy network and evaluation network included in the strategy-evaluation network; generating trading actions through the strategy network; determining the value assessment of the trading actions through the evaluation network; and updating the parameters of the strategy network and evaluation network through backpropagation of network gradients based on the reward function and value assessment. In the example, the output of the strategy-evaluation network includes the mean and standard deviation of the declared electricity price and declared electricity volume, as well as the probability distribution of the trading actions determined based on the mean and standard deviation.

[0050] For example, the attention mechanism, as the first layer of the network, dynamically assigns weights to different parts of the original state vector to enhance key information. (dimension is) Through three different linear layers (weight matrix) W Q’ , W K , W V Projecting the query vector yields the query vector. Q’ Key vector K Sum value vector V : Q=s W Q’ , K=sW K , V=sW V .

[0051] The formula for calculating attention is: .

[0052] The policy network is a policy function whose input is the state feature vector processed by the attention module. The output is a continuous transaction action (including the declared electricity price and the declared electricity quantity), and may include the mean and standard deviation of the declared electricity price and electricity quantity. And the probability distribution of trading actions composed of these parameters. ,in, This is the average electricity price. This is the average amount of electricity consumed. For the standard deviation of electricity prices, This represents the standard deviation of the electricity consumption.

[0053] In the example, mean squared error is used as the loss function during backpropagation of network gradients. In the example, both the policy network and the evaluation network consist of three fully connected hidden layers. However, this disclosure is not limited to this; the policy network and evaluation network can be constructed with more or fewer layers and different numbers of units per layer.

[0054] For example, both the policy network and the evaluation network are designed with three fully connected hidden layers, with the number of units in the hidden layers being 256-128-32, and an attention mechanism module is added to the input of the hidden layers.

[0055] Policy network: Input is a state vector processed by the attention mechanism. After network processing, the probability distribution of output actions is obtained. .

[0056] Evaluating the network: Input vectors one by one and actions from the policy network Output a scalar And the reward value, which represents the value assessment of the current state. When the action distribution... After all the data is input, based on each action... Choose the action with the highest reward value.

[0057] The calculation method for network forward propagation is as follows: Weighted State The system passes through three fully connected hidden layers of the policy network in sequence, with each layer typically followed by a ReLU activation function to introduce nonlinearity.

[0058]

[0059]

[0060]

[0061] The formula for calculating the ReLU function is as follows: The matrix W is a randomly generated parameter matrix, and its specific values ​​are determined by the later model training.

[0062] Output of the policy network As shown below: .

[0063] The softmax function outputs a probability vector, and the output of the last hidden layer of the policy network is a 32-dimensional vector. The softmax function then outputs a probability vector. The calculation formula is as follows: .

[0064] Will The input is fed into the evaluation network to obtain an evaluation score. The hidden layers of the evaluation network are similar to those of the policy network, and its output layer is a linear transformation layer that maps the last hidden unit to a scalar value. : .

[0065] Then, select The highest-level strategy will include the declared electricity volume and declared electricity price for the predetermined time period. Then, the agent model parameters are updated through backpropagation of network gradients. The formula for calculating the backpropagation gradient is as follows. Figure 2 As shown, the policy-evaluation network uses mean squared error as the loss function to make... Minimize, that is, .

[0066] The gradient update formula for evaluating the network is: .

[0067] The intelligent trading method for the electricity market according to exemplary embodiments of this disclosure can improve decision-making efficiency and accuracy. By approximating complex market state mappings through deep neural networks, it overcomes the dependence of traditional methods on model simplification and achieves near-optimal strategies in environments with incomplete information. It enhances market adaptability by possessing online learning capabilities, dynamically adjusting bidding strategies based on market price fluctuations and policy changes, adapting to market uncertainties in scenarios with a high proportion of renewable energy. It optimizes returns by not pursuing absolute accuracy in single price predictions, but rather considering both risk and long-term effects to maximize long-term cumulative returns, thus better adapting to the characteristics of large daily rolling market fluctuations and complex market information. Furthermore, it improves computational efficiency by employing a distributed training architecture, supporting parallel processing of large-scale historical data, and solving the curse of dimensionality problem in traditional game theory methods.

[0068] A photovoltaic power station with an installed capacity of 50 MWh was selected as the target site for testing. Daily rolling transactions were conducted on the second, third, and fourth days, with the entire test lasting one month. The final benefit statistics are shown in Table 1 below.

[0069] Table 1

[0070] The statistics above show that the smart trading method for the electricity market according to the exemplary embodiments of this disclosure achieved positive revenue growth in the daily rolling trading monthly test at the power station on the second, third, and fourth days.

[0071] Figure 3 This is a block diagram illustrating an intelligent trading device for an electricity market according to an exemplary embodiment of the present disclosure.

[0072] like Figure 3 As shown, an exemplary embodiment of the present disclosure provides an intelligent trading device 300 for the electricity market, comprising: a market state feature vector determination unit 301, configured to acquire historical trading data of the electricity market, power grid physical state data, wind farm data, and external environment data, and determine a market state feature vector based on the acquired data; and a trading action generation unit 302, configured to generate trading actions for a predetermined time period based on the market state feature vector through a strategy-evaluation network incorporating an attention mechanism, wherein the trading actions include submitting electricity volume and submitting electricity price, and the reward function applied during the training process of the strategy-evaluation network includes a core benefit term, a deviation penalty term, a risk penalty term, and a constraint reward term.

[0073] In the example, the market state feature vector determination unit 301 is further configured to: select a predetermined number of historical transaction data from the historical transaction data of the electricity market based on Euclidean distance, whose timestamps are closest to the current transaction date, to form a sample set; determine key features based on the sample set, the key features including at least the supply and demand tension index, price fluctuation statistics, and time series features; and determine the market state feature vector based on the key features and real-time observed electricity market transaction data, power grid physical state data, and wind farm data.

[0074] In the example, the output of the strategy-evaluation network includes the mean and standard deviation of the declared electricity price and declared electricity volume, as well as the probability distribution of the transaction action determined based on the mean and standard deviation.

[0075] In the example, the core revenue item is determined based on the declared electricity price, the cleared electricity price, and the cleared volume; the deviation penalty item is determined based on the difference between the declared volume and the cleared volume; and the risk penalty item is determined based on the value at risk at a preset confidence level and the current expected rate of return, calculated from historical revenue data of similar days.

[0076] In the example, the training process of the policy-evaluation network includes: mapping the market state feature vector to query vector, key vector, and value vector through an attention layer, and determining the weighted feature vector based on the query vector, key vector, and value vector through a normalized exponential function; inputting the weighted feature vector into the policy network and evaluation network included in the policy-evaluation network; generating trading actions through the policy network; determining the value assessment of the trading actions through the evaluation network; and updating the parameters of the policy network and evaluation network through backpropagation of network gradients based on the reward function and value assessment.

[0077] In the example, mean squared error is used as the loss function during the backpropagation of network gradients.

[0078] In the example, both the policy network and the evaluation network include three fully connected hidden layers.

[0079] The above combination Figure 1 and Figure 2 The specific operations shown are respectively by Figure 3The corresponding units in the intelligent trading device 300 for the electricity market shown here will execute the above actions; specific operational details will not be elaborated here. The intelligent trading device for the electricity market according to the exemplary embodiments of this disclosure can improve decision-making efficiency and accuracy. By approximating complex market state mappings through deep neural networks, it overcomes the dependence of traditional methods on model simplification and achieves near-optimal strategies in environments with incomplete information. It enhances market adaptability by possessing online learning capabilities, dynamically adjusting bidding strategies based on market price fluctuations and policy changes, adapting to market uncertainties in scenarios with a high proportion of renewable energy. It optimizes returns by not pursuing absolute accuracy in single price predictions, but rather considering both risk and long-term effects to maximize long-term cumulative returns, better adapting to the characteristics of large daily rolling market fluctuations and complex market information. Furthermore, it improves computational efficiency by adopting a distributed training architecture, supporting parallel processing of large-scale historical data, and solving the curse of dimensionality problem in traditional game theory methods.

[0080] Figure 4 This is a block diagram illustrating a computing system including at least one computing device and at least one storage device of storage instructions according to an exemplary embodiment of the present disclosure.

[0081] like Figure 4 As shown, the computing system 400 provided according to an exemplary embodiment of the present invention includes a computing device 401 and a storage device 402. The storage device 402 stores computer-executable instructions. When the computer-executable instructions are executed by the computing device 401, the smart trading method for the electricity market described in any of the foregoing embodiments is executed.

[0082] The computing device 401 can be deployed in a server or client, or on a node device in a distributed network environment. Furthermore, the computing device 401 can be a PC, tablet, personal digital assistant, smartphone, web application, or other device capable of executing the aforementioned set of instructions. Here, the computing device is not necessarily a single computing device; it can be any collection of devices or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. The computing device can also be part of an integrated control system or system manager, or can be configured to interconnect with a portable electronic device locally or remotely (e.g., via wireless transmission). In the computing device, the processor includes a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, the processor also includes analog processors, digital processors, microprocessors, multi-core processors, processor arrays, network processors, etc.

[0083] According to another aspect of this disclosure, a computer-readable storage medium is provided that stores instructions, which, when executed by at least one computing device, cause the at least one computing device to perform the smart trading method for an electricity market as described in any of the foregoing embodiments. The computer-readable storage medium includes magnetic media such as floppy disks and magnetic tapes, optical media (including optical disc (CD) ROMs and DVD ROMs), magneto-optical media such as floppy discs, hardware devices such as ROMs and RAMs designed for storing and executing program commands, and flash memory. The instructions may include language code executable by a computer using an interpreter and machine language code generated by a compiler.

[0084] By adopting this disclosure, decision-making efficiency can be improved, enabling 24 / 7 monitoring of multi-source market data (e.g., prices, load, renewable energy output, etc.) and instantaneous completion of the entire process from analysis to decision-making, seizing short-term arbitrage opportunities that are difficult to capture manually; decision-making accuracy can be improved by approximating complex market state mappings through deep neural networks, overcoming the dependence of traditional methods on model simplification, and achieving near-optimal strategies in environments with incomplete information; market adaptability can be enhanced by having online learning capabilities, dynamically adjusting pricing strategies according to market price fluctuations and policy changes, adapting to the market uncertainty in scenarios with a high proportion of renewable energy; returns can be optimized by not pursuing absolute accuracy in single price predictions, but rather by comprehensively considering risks and long-term effects to maximize the optimal trading strategy for long-term cumulative returns, better adapting to the characteristics of large daily rolling market fluctuations and complex market information; and computational efficiency can be improved by adopting a distributed training architecture, supporting parallel processing of large-scale historical data, and solving the curse of dimensionality problem in traditional game theory methods.

[0085] The processes, methods, or algorithms disclosed herein can be transmitted to, or implemented by, a processing device, controller, or computer, which may include any existing programmable electronic control unit or a dedicated electronic control unit. Similarly, the processes, methods, or algorithms can be stored in various forms as data and instructions executable by a controller or computer, including but not limited to information permanently stored on non-writable storage media (such as ROM devices) and information variablely stored on writable storage media (such as floppy disks, magnetic tapes, CDs, RAM devices, and other magnetic and optical media). The processes, methods, or algorithms can also be implemented in a software executable object. Optionally, the processes, methods, or algorithms can be implemented wholly or partially using suitable hardware components (such as ASICs, FPGAs, state machines, controllers, or other hardware components or devices) or a combination of hardware components, software components, and firmware components.

[0086] Although this disclosure includes specific examples, it will be apparent to those skilled in the art that various changes in form and detail may be made to these examples without departing from the spirit and scope of the claims and their equivalents. The examples described herein are to be considered merely for descriptive purposes and not for limiting purposes. The description of features or aspects in each example is to be considered applicable to similar features or aspects in other examples. Suitable results may be obtained if the described techniques are performed in a different order, and / or if components in the described system, architecture, apparatus, or circuit are combined in a different manner and / or if components in the described system, architecture, apparatus, or circuit are replaced or supplemented with other components or their equivalents. Therefore, the scope of this disclosure is not limited by the specific embodiments but by the claims and their equivalents, and all variations within the scope of the claims and their equivalents shall be construed as included in this disclosure.

Claims

1. A smart trading method for the electricity market, characterized in that, The intelligent trading method includes: Acquire historical transaction data of the electricity market, power grid physical status data, wind farm data, and external environment data, and determine market status feature vectors based on the acquired data; Based on the market state feature vector, a strategy-evaluation network incorporating an attention mechanism generates trading actions for a predetermined time period. The transaction actions include submitting electricity volume and submitting electricity price. The reward function applied during the training of the strategy-evaluation network includes a core revenue term, a deviation penalty term, a risk penalty term, and a constraint reward term.

2. The smart trading method for the electricity market according to claim 1, characterized in that, The steps for determining market state feature vectors based on the acquired data include: Based on Euclidean distance, a predetermined number of historical transaction data points with timestamps closest to the current transaction date are selected from the historical transaction data of the electricity market to form a sample set; Based on the sample set, key features are determined, including at least the supply and demand tension index, price fluctuation statistics, and time series features. Based on the aforementioned key features and real-time observed electricity market transaction data, power grid physical status data, and wind farm data, the market status feature vector is determined.

3. The smart trading method for the electricity market according to claim 1, characterized in that, The output of the strategy-evaluation network includes the mean and standard deviation of the declared electricity price and declared electricity volume, as well as the probability distribution of the transaction action determined based on the mean and standard deviation.

4. The intelligent trading method for the electricity market according to claim 1, characterized in that, The core revenue item is determined based on the declared electricity price, the cleared electricity price, and the cleared volume. The deviation penalty item is determined based on the difference between the declared volume and the cleared volume. The risk penalty item is determined based on the value at risk at a preset confidence level and the current expected rate of return, calculated from historical revenue data of similar days.

5. The intelligent trading method for the electricity market according to claim 1, characterized in that, The training process of the policy-evaluation network includes: The market state feature vector is mapped to query vector, key vector and value vector through the attention layer, and the weighted feature vector is determined based on the query vector, key vector and value vector through the normalized exponential function; The weighted feature vectors are respectively input into the policy network and the evaluation network included in the policy-evaluation network; The transaction action is generated through the strategy network; The value assessment of the transaction is determined through the evaluation network. Based on the reward function and the value assessment, the parameters of the policy network and the evaluation network are updated through backpropagation of network gradients.

6. The smart trading method for the electricity market according to claim 5, characterized in that, Mean squared error is used as the loss function during the backpropagation of the network gradient.

7. The smart trading method for the electricity market according to claim 5, characterized in that, Both the policy network and the evaluation network include three fully connected hidden layers.

8. A smart trading device for the electricity market, characterized in that, The intelligent trading device includes: The market state feature vector determination unit is configured to acquire historical transaction data of the electricity market, power grid physical state data, wind farm data, and external environment data, and determine the market state feature vector based on the acquired data; The trading action generation unit is configured to generate trading actions for a predetermined time period based on the market state feature vector, using a strategy-evaluation network incorporating an attention mechanism. The transaction actions include submitting electricity volume and submitting electricity price. The reward function applied during the training of the strategy-evaluation network includes a core revenue term, a deviation penalty term, a risk penalty term, and a constraint reward term.

9. A computing system comprising at least one computing device and at least one storage device for storing instructions, characterized in that, When the instruction is executed by the at least one computing device, it causes the at least one computing device to perform the smart trading method for the electricity market according to any one of claims 1 to 7.

10. A computer-readable storage medium for storing instructions, characterized in that, When the instruction is executed by at least one computing device, it causes the at least one computing device to perform the smart trading method for the electricity market according to any one of claims 1 to 7.

Images