Method and system for predicting quantity report, quotation and bidding by building load aggregator

Abstract

The invention provides a prediction method and system for quantity report, quotation and bidding by a building load aggregator, and the method comprises the steps: collecting bidding historical data, and carrying out the preprocessing of the bidding historical data; inputting the preprocessed bidding historical data as an input item into a pre-trained quantity and quotation behavior prediction model to obtain a bidding action probability distribution result of the competitor; and based on the bidding action probability distribution result of the competitor, in a preset bidding decision simulation environment, predicting an optimal quantity report quotation bidding strategy of our party as a building load aggregator, and outputting a final prediction result. According to the method, the market competition pattern can be dynamically predicted, and the optimal quantity and quotation strategy is generated, so that the bid winning rate and the economic benefit of the building load aggregator in the demand response real-time market are improved.

Description

Technical Field

[0001] This invention relates to the field of quantity, price and bid prediction technology, and in particular to a method and system for predicting quantity, price and bid for building load aggregators. Background Technology

[0002] Building load aggregators, as crucial intermediaries connecting a vast number of dispersed building users with the electricity market, participate in market bidding as a single entity by integrating and managing their adjustable load resources. Establishing a scientific and robust bidding decision-making methodology is of paramount importance for enhancing the market competitiveness of building load aggregators.

[0003] Analysis of existing technologies reveals that bidding decisions for aggregators typically follow a linear, two-stage process: The first stage focuses on the aggregator itself, using deterministic forecasting or basic probability models to estimate its potential to stably reduce load during future response periods; the second stage, based on this load level, calculates the incentive costs payable to users and, according to the cost-plus principle (adding a fixed profit margin above the cost), constructs the final bid price. However, when applied to a highly competitive real-time electricity market comprised of multiple rational actors, these existing technologies suffer from the following shortcomings: their decision-making framework is non-game-theoretic, failing to include quantitative analysis and prediction of the bidding behavior of other competing market participants. The clearing price of the electricity market and an individual's ability to win a bid are not determined by a single bidder, but rather by the supply and demand curves formed by the collective bidding of all market participants.

[0004] In conclusion, while existing technologies provide a basic reference for aggregators to participate in the market, their decision-making logic and neglect of market competition dynamics make them unable to meet the business goal of achieving long-term and stable profitability in the modern real-time electricity market. Summary of the Invention

[0005] The purpose of this invention is to propose a method and system for predicting the bidding price of building load aggregators, and to solve the technical problem of how to achieve the globally optimal bidding strategy with the goal of maximizing expected revenue.

[0006] On the one hand, a method for predicting the bidding price of building load aggregators is provided, including:

[0007] Collect historical bidding data and preprocess the historical bidding data;

[0008] The preprocessed bidding history data is used as input to a pre-trained volume and price quotation behavior prediction model to obtain the probability distribution of competitors' bidding actions. The pre-trained volume and price quotation behavior prediction model is based on historical data, takes influencing factors as input, takes the discretized actions actually taken by competitors as labels, performs supervised learning training on a preset neural network model, and predicts the probability distribution of competitors' bidding actions based on the input factors.

[0009] Based on the probability distribution of the competitors' bidding actions, in a preset bidding decision simulation environment, the optimal quantity and price bidding strategy of our party as a building load aggregator is predicted, and the final prediction result is output.

[0010] Preferably, the bidding history data includes at least:

[0011] Historical market data, including real-time market clearing prices and total system demand at fixed time intervals;

[0012] Competitors' historical bidding data, including the aggregator's historical bid volume and corresponding price data;

[0013] Our historical bidding data includes the historical bids, prices, and winning results of our aggregators;

[0014] External factor data, including daily maximum / minimum temperature, weather type, and whether it is a holiday / workday.

[0015] Preferably, the preprocessing of the bidding history data includes at least discretizing each continuous variable in the bidding history data and outputting a discretized format including preset parameter types and typical values.

[0016] Preferably, the reporting and pricing behavior prediction model includes at least a plurality of hidden layers and an output layer; wherein each hidden layer has at least 128 neurons and ReLU is set as the activation function; the output layer is set with a Softmax function, and its output is a probability distribution with a sum of one.

[0017] Preferably, the preset bidding decision simulation environment includes at least the following: when the agent representing our building load aggregator performs a bidding action, it randomly samples and generates one or more competitors' bidding actions based on the current input features and the probability distribution results of competitors' bidding actions.

[0018] Combining our actions with those of the simulated opponent, the system calculates the clearing result according to the preset real-time market clearing rules and derives a reward value related to our economic gains.

[0019] Based on the reward value, through continuous iteration and strategy updates, until the preset iteration target is reached, the optimal bidding strategy for quantity quotation is output.

[0020] Preferably, the random sampling to generate bids from one or more competitors includes,

[0021] Select a specific bidding action based on the current market conditions. This bidding action consists of a quantity level and a price level.

[0022] Upon receiving a bid action, the bidding behavior prediction model is invoked to generate a probability distribution of actions for each competitor based on the current market state. Random sampling is then performed based on this distribution to obtain the bidding actions of all simulated competitors.

[0023] Preferably, the step of calculating the clearing result according to the preset real-time market clearing rules includes:

[0024] All bids are sorted from lowest to highest price and accepted sequentially until the system's total demand is met. The final settlement price for all winning bids is then output, and the bid for the last winning bid is set as the marginal price.

[0025] Preferably, the process of deriving a reward value related to our economic gains includes:

[0026] The corresponding reward value is output through a preset reward function, wherein the reward function is to multiply the difference between the clearing price and our bid by the amount of electricity we won the bid and output the reward value. If we fail to win the bid, the reward value is set to zero.

[0027] Preferably, the step of continuously iterating and updating the strategy based on the reward value includes:

[0028] In each iteration, the agent selects the known optimal action with a probability of 1-∈ based on the current Q-value table, and randomly selects an action with a probability of ∈ to explore.

[0029] After performing the bidding action, the agent receives a corresponding reward from the environment and updates the Q-value table based on this reward;

[0030] Through continuous iteration and strategy updates, the Q value gradually converges until it reflects the long-term expected return of performing a bidding action under a specific state.

[0031] Once the training converges, the obtained Q-value table or the trained policy network will be output as our optimal quantity quotation bidding strategy.

[0032] On the other hand, a prediction system for building load aggregators' bid pricing and quantity reporting is also provided, to implement the aforementioned prediction method for building load aggregators' bid pricing and quantity reporting, including:

[0033] The data processing module is used to collect historical bidding data and preprocess the historical bidding data.

[0034] The probability calculation module is used to input the preprocessed bidding history data as input to the pre-trained bidding behavior prediction model to obtain the probability distribution of competitors' bidding actions. The pre-trained bidding behavior prediction model is based on historical data, takes influencing factors as input, takes the discretized actions actually taken by competitors as labels, performs supervised learning training on the preset neural network model, and predicts the probability distribution of competitors' bidding actions based on the input factors.

[0035] The simulation module is used to predict the optimal bidding strategy for our company as a building load aggregator based on the probability distribution of the competitors' bidding actions in a preset bidding decision simulation environment, and outputs the final prediction result.

[0036] In summary, implementing the embodiments of the present invention has the following beneficial effects:

[0037] The present invention provides a method and system for predicting the bidding and pricing of building load aggregators, which is applied to the bidding decision optimization of building load aggregators. By constructing a competitor behavior prediction model and a bidding strategy optimization method based on reinforcement learning, aggregators can dynamically predict the market competition pattern and generate the optimal bidding and pricing strategy, thereby improving the success rate and economic benefits of building load aggregators in the real-time demand response market. It is applicable to scenarios with adjustable loads, such as large-scale commercial buildings, industrial parks, or distributed energy systems. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0039] Figure 1 This is a schematic diagram of the main process of a method for predicting the bidding price of building load aggregators in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0041] like Figure 1 The diagram shown is an embodiment of a method for predicting quantity and price bids by a building load aggregator, provided by the present invention. In this embodiment, the method includes the following steps:

[0042] Step S1: Collect historical bidding data and preprocess the historical bidding data;

[0043] Step S2: Input the preprocessed bidding history data as input to the pre-trained volume and price quotation behavior prediction model to obtain the probability distribution of the competitor's bidding actions; wherein, the pre-trained volume and price quotation behavior prediction model is based on historical data, takes influencing factors as input, takes the discretized actions actually taken by the competitor as labels, performs supervised learning training on the preset neural network model, and predicts the probability distribution of the competitor's bidding actions based on the input factors.

[0044] Step S3: Based on the probability distribution of the competitors' bidding actions, predict the optimal quantity and price bidding strategy of our company as a building load aggregator in a preset bidding decision simulation environment, and output the final prediction result.

[0045] In one specific embodiment of the present invention, step S1 aims to transform raw, continuous market and environmental data into a discretized format that the model can process.

[0046] First, collect detailed historical data from the past 24 months, including:

[0047] 1. Historical market data: obtained from electricity market operators, including real-time market clearing price (RTP) every 15 minutes and total system demand.

[0048] 2. Competitors' historical bidding data: Historical bid volume and price data of aggregators A, B, and C.

[0049] 3. Our historical bidding data: Our (aggregator D) historical bid volume, price quotation and winning bid status.

[0050] 4. External factor data: including daily maximum / minimum temperature, weather type (sunny, cloudy, rainy, snowy), and whether it is a holiday / working day.

[0051] Secondly, the continuous variables in the bidding history data are discretized, and the output is a discretized format including preset parameter types and typical values. K-Means clustering algorithm is usually used to discretize the continuous variables. The parameter settings and typical values ​​are as follows.

[0052] variable Number of clusters K (typical value) Quotation 10 Reporting volume 10 Market clearing price 10 System General Requirements 10 Weather temperature 5

[0053] Step S2 aims to create a probabilistic model that can predict competitors' bidding behavior based on the current market state. A Multilayer Perceptron (MLP) neural network is used to construct the predictive model. The bidding behavior prediction model includes at least multiple hidden layers and an output layer; each hidden layer has at least 128 neurons and uses ReLU as the activation function; the output layer uses a Softmax function, and its output is a probability distribution with a sum of 1. Specifically, a typical architecture includes three hidden layers, each with 128 neurons, using ReLU as the activation function, while the output layer uses a Softmax function to ensure that the output is a probability distribution with a sum of 1. The input features of this model are combinations of market states after discretization in Step 1, such as [time point, weather type, holiday type, total system demand level, historical market price level]; its output is a probability distribution vector. If there are 10 levels for both bidding volume and price, the dimension of the output vector is 100. During the model training phase, we use historical data processed in Step 1, taking the market state as input and the discretized actions actually chosen by competitors in that state as labels. The training process uses the Adam optimizer with a typical learning rate of 0.001, a batch size of 64, and a total of 100 training epochs. After training, the model can generate all possible bidding actions of competitors and their corresponding probabilities based on a given market state.

[0054] One embodiment establishes a neural network-based model for predicting competitor bidding behavior, specifically including: Model Input: The discretized historical factors influencing competitor bidding behavior, processed in step one, are used as input features of the model. Model Output: The output layer of the model is designed to map the probability distribution of different discretized bidding actions taken by competitors under the current input features. Model Training: Using historical data, with influencing factors as input and the actual discretized actions taken by competitors as labels, the neural network model is trained under supervised learning until the model can accurately predict the probability distribution of competitors' bidding actions based on the input factors.

[0055] Step S3 uses the pre-trained opponent prediction model to create a virtual market environment for our intelligent agent to learn and experiment.

[0056] Specifically, the preset bidding decision simulation environment includes at least the following steps: when the agent representing our building load aggregator performs a bidding action, it randomly samples and generates one or more competitors' bidding actions based on the current input features and the probability distribution of competitors' bidding actions; combining our actions and the simulated competitors' actions, it calculates the clearing result according to preset real-time market clearing rules and obtains a reward value related to our economic benefits; based on the reward value, it continuously iterates and updates the strategy until a preset iteration target is reached, and then outputs the optimal quantity and price bidding strategy.

[0057] The random sampling to generate one or more competitors' bidding actions includes: selecting a specific bidding action based on the current market state, which consists of a quantity level and a price level; after receiving the bidding action, calling the quantity and price behavior prediction model to generate the probability distribution of the action for each competitor based on the current market state, and performing random sampling based on this distribution to obtain all simulated competitors' bidding actions.

[0058] The calculation of the clearing result according to the preset real-time market clearing rules includes sorting all bids from low to high and accepting them in sequence until the total system demand is met, outputting the final settlement price of all winning bids, and setting the bid of the last winning bid as the marginal price.

[0059] The process of obtaining a reward value related to our economic gains includes outputting the corresponding reward value through a preset reward function. The reward function is to multiply the difference between the clearing price and our bid by the amount of electricity we won the bid, and output the reward value as the reward value if we fail to win the bid.

[0060] The step of continuously iterating and updating the policy based on the reward value includes the following steps: In each iteration, the agent selects the known optimal action with a probability of 1-∈ based on the current Q-value table, and randomly selects an action for exploration with a probability of ∈; After executing the bidding action, the agent obtains the corresponding reward from the environment and updates the Q-value table based on this reward; Through continuous iteration and policy updates, the Q-value gradually converges until it reflects the long-term expected return of executing a certain bidding action in a specific state; When the training converges, the obtained Q-value table or the trained policy network is output as our optimal bidding strategy.

[0061] It should be noted that its operating mechanism is as follows: Our agent selects a specific bidding action based on the current market state. This action consists of a quantity level and a price level. Upon receiving this action, the simulation environment immediately invokes the competitor prediction model. This model generates a probability distribution for each competitor's action based on the current state and performs random sampling based on this distribution to obtain the bidding actions of all simulated competitors. Subsequently, the environment's built-in simplified unified clearing pricing model begins operation. All bids are sorted from lowest to highest price and accepted sequentially until the system's total demand is met. The final settlement price for all winning units will be equal to the price of the last winning unit, i.e., the marginal price.

[0062] To guide the agent's learning direction, a reward function is designed. This function calculates our profit based on the market clearing result and uses it as the reward signal. The specific formula is: R = (Clearing Price - Our Bid) × Our Winning Electricity Volume. If we fail to win the bid, the reward R for that round is 0.

[0063] In this simulation environment, the Epsilon(∈) greedy algorithm based on Q-Learning is used to train our agent, aiming to find a bidding strategy that maximizes long-term cumulative returns. The core components of this algorithm include: an agent representing our building load aggregator; an action space consisting of 100 discretized combinations of supply and demand; and a large Q-table. The index of this Q-table is a specific "market state," the columns are "all possible actions," and each value Q(s,a) in the table represents the expected long-term return that can be obtained by performing action a in state s.

[0064] The training process is accomplished through numerous iterations, typically set at 50,000, but ranging from 10,000 to 100,000. An Epsilon(∈) greedy strategy is used, with ∈ initially set to 1.0, indicating completely random exploration in the early stages of training. Subsequently, the value of ∈ decays with each iteration, typically by a decay factor of 0.9995, until a predetermined minimum value is reached, usually 0.05, ranging from 0.01 to 0.1.

[0065] The update of the Q-value is regulated by two key parameters: the learning rate (α) and the discount factor (γ). The learning rate determines the extent to which newly acquired information overwrites old information; its typical value is 0.1, with a range of 0.01 to 0.2. The discount factor measures the importance of future rewards relative to current rewards; its typical value is 0.95, with a range of 0.9 to 0.99.

[0066] After training, the final converged Q-value table constitutes the optimal bidding strategy. In actual bidding decision-making, the process is very clear: First, obtain and discretize the current real-time market state; next, find the row corresponding to that state in the Q-value table; finally, select the action with the maximum Q-value in that row (i.e., the combination of quantity and price), and use this as the optimal bidding decision for this instance. In this way, the complex market game problem is transformed into a solvable machine learning problem, providing building load aggregators with a data-driven, adaptive, and continuously optimizing intelligent bidding decision-making tool.

[0067] This invention overcomes the shortcomings of existing technologies that rely on subjective experience and ignore market competition, significantly improving decision-making efficiency and economic benefits. Its beneficial effects include: introducing a game theory perspective to increase the success rate of bidding: through a newly added competitor behavior prediction model, decision-making is transformed from passive cost accounting to active dynamic game theory. Achieving automatic optimization and increasing profits: using reinforcement learning algorithms to replace subjective judgment, it can automatically solve for the optimal bidding strategy that maximizes expected returns. Dynamic strategy adaptation and enhanced robustness: decisions can be adjusted according to real-time market changes, the strategy is no longer rigid, and it can flexibly respond to market fluctuations.

[0068] Embodiments of the present invention also provide a prediction system for building load aggregators' bid pricing and quantity estimates, for implementing the aforementioned prediction method for building load aggregators' bid pricing and quantity estimates, comprising:

[0069] The data processing module is used to collect historical bidding data and preprocess the historical bidding data.

[0070] The probability calculation module is used to input the preprocessed bidding history data as input to the pre-trained bidding behavior prediction model to obtain the probability distribution of competitors' bidding actions. The pre-trained bidding behavior prediction model is based on historical data, takes influencing factors as input, takes the discretized actions actually taken by competitors as labels, performs supervised learning training on the preset neural network model, and predicts the probability distribution of competitors' bidding actions based on the input factors.

[0071] The simulation module is used to predict the optimal bidding strategy for our company as a building load aggregator based on the probability distribution of the competitors' bidding actions in a preset bidding decision simulation environment, and outputs the final prediction result.

[0072] It should be noted that the system described in the above embodiments corresponds to the method described in the above embodiments. Therefore, the parts of the system described in the above embodiments that are not described in detail can be obtained by referring to the content of the method described in the above embodiments, and will not be repeated here.

[0073] In summary, implementing the embodiments of the present invention has the following beneficial effects:

[0074] The present invention provides a method and system for predicting the bidding and pricing of building load aggregators, which is applied to the bidding decision optimization of building load aggregators. By constructing a competitor behavior prediction model and a bidding strategy optimization method based on reinforcement learning, aggregators can dynamically predict the market competition pattern and generate the optimal bidding and pricing strategy, thereby improving the success rate and economic benefits of building load aggregators in the real-time demand response market. It is applicable to scenarios with adjustable loads, such as large-scale commercial buildings, industrial parks, or distributed energy systems.

[0075] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for predicting the quantity, price, and bidding of building load aggregators, characterized in that, include: Collect historical bidding data and preprocess the historical bidding data; The preprocessed bidding history data is used as input to a pre-trained volume and price quotation behavior prediction model to obtain the probability distribution of competitors' bidding actions. The pre-trained volume and price quotation behavior prediction model is based on historical data, takes influencing factors as input, takes the discretized actions actually taken by competitors as labels, performs supervised learning training on a preset neural network model, and predicts the probability distribution of competitors' bidding actions based on the input factors. Based on the probability distribution of the competitors' bidding actions, in a preset bidding decision simulation environment, the optimal quantity and price bidding strategy of our party as a building load aggregator is predicted, and the final prediction result is output.

2. The method as described in claim 1, characterized in that, The bidding history data includes at least the following: Historical market data, including real-time market clearing prices and total system demand at fixed time intervals; Competitors' historical bidding data, including the aggregator's historical bid volume and corresponding price data; Our historical bidding data includes the historical bids, prices, and winning results of our aggregators; External factor data, including daily maximum / minimum temperature, weather type, and whether it is a holiday / workday.

3. The method as described in claim 2, characterized in that, The preprocessing of the bidding history data includes at least discretizing each continuous variable in the bidding history data and outputting a discretized format including preset parameter types and typical values.

4. The method as described in claim 1, characterized in that, The reported volume and price quotation behavior prediction model includes at least multiple hidden layers and an output layer; wherein each hidden layer has at least 128 neurons and ReLU is set as the activation function; the output layer is set with a Softmax function, and its output is a probability distribution with a sum of one.

5. The method as described in claim 1, characterized in that, The preset bidding decision simulation environment includes at least the following: when the intelligent agent representing our building load aggregator performs a bidding action, it randomly samples and generates one or more competitors' bidding actions based on the current input features and the probability distribution results of competitors' bidding actions. Combining our actions with those of the simulated opponent, the system calculates the clearing result according to the preset real-time market clearing rules and derives a reward value related to our economic gains. Based on the reward value, through continuous iteration and strategy updates, until the preset iteration target is reached, the optimal bidding strategy for quantity quotation is output.

6. The method as described in claim 5, characterized in that, The action of generating one or more competitors' bids through random sampling includes, Select a specific bidding action based on the current market conditions. This bidding action consists of a quantity level and a price level. Upon receiving a bid action, the bidding behavior prediction model is invoked to generate a probability distribution of actions for each competitor based on the current market state. Random sampling is then performed based on this distribution to obtain the bidding actions of all simulated competitors.

7. The method as described in claim 6, characterized in that, The calculation of the clearing result based on the preset real-time market clearing rules includes, All bids are sorted from lowest to highest price and accepted sequentially until the system's total demand is met. The final settlement price for all winning bids is then output, and the bid for the last winning bid is set as the marginal price.

8. The method as described in claim 7, characterized in that, The process of deriving a reward value related to our economic gains includes, The corresponding reward value is output through a preset reward function, wherein the reward function is to multiply the difference between the clearing price and our bid by the amount of electricity we won the bid and output the reward value. If we fail to win the bid, the reward value is set to zero.

9. The method as described in claim 8, characterized in that, The process of continuously iterating and updating the strategy based on the reward value includes, In each iteration, the agent selects the known optimal action with a probability of 1-∈ based on the current Q-value table, and randomly selects an action with a probability of ∈ to explore. After performing the bidding action, the agent receives a corresponding reward from the environment and updates the Q-value table based on this reward; Through continuous iteration and strategy updates, the Q value gradually converges until it reflects the long-term expected return of performing a bidding action under a specific state. Once the training converges, the obtained Q-value table or the trained policy network will be output as our optimal quantity quotation bidding strategy.

10. A prediction system for building load aggregators' reported quantity and price bids, used to implement the method as described in any one of claims 1-9, characterized in that, include: The data processing module is used to collect historical bidding data and preprocess the historical bidding data. The probability calculation module is used to input the preprocessed bidding history data as input to the pre-trained bidding behavior prediction model to obtain the probability distribution of competitors' bidding actions. The pre-trained bidding behavior prediction model is based on historical data, takes influencing factors as input, takes the discretized actions actually taken by competitors as labels, performs supervised learning training on the preset neural network model, and predicts the probability distribution of competitors' bidding actions based on the input factors. The simulation module is used to predict the optimal bidding strategy for our company as a building load aggregator based on the probability distribution of the competitors' bidding actions in a preset bidding decision simulation environment, and outputs the final prediction result.

Images