Electric power spot day-ahead market auxiliary quotation method

Abstract

The invention discloses an electric power spot day-ahead market auxiliary quotation method, and relates to the technical field of electric power system automation, and the method comprises the steps: collecting power grid operation, weather, market and carbon market data through multi-source data fusion preprocessing, and carrying out the processing through a space-time generative adversarial network and a natural language processing model, so as to form a multi-dimensional data set; embedding a power grid power balance constraint by using a graph neural network to carry out topology modeling, and extracting spatio-temporal characteristics by combining a time sequence neural network and an attention mechanism; and a double-layer optimization model is constructed to realize collaborative decision-making of futures and spot markets, and carbon-electricity joint optimization and a dynamic shipping space adjustment mechanism are synchronously integrated. According to the method, the problems of data isolation, inaccurate prediction, insufficient risk control and the like of a traditional quotation method are solved, the quotation precision and the market competitiveness are improved, an auxiliary quotation scheme giving consideration to benefits and risks is provided for power generation enterprises, and the method is suitable for quotation strategy optimization of the current market of electric power spot goods.

Description

Technical Field

[0001] The present invention relates to the technical field of power system automation, and in particular to an auxiliary quotation method for the electricity spot day-ahead market. Background Art

[0002] Against the backdrop of deepening power system reform and the gradual advancement of the electricity spot market, traditional power market quotation models face numerous challenges. Existing power market quotation methods primarily rely on marginal cost pricing, historical data statistics, and trader experience and judgment. These methods suffer from data isolation, insufficient forecasting accuracy, and weak risk control.

[0003] Specifically, on the one hand, electricity market data is scattered across multiple independent systems, such as grid operation monitoring systems, weather forecasting systems, and market trading platforms. This creates data silos, making effective integration and comprehensive utilization difficult, resulting in a lack of comprehensive data support for bidding decisions. On the other hand, traditional forecasting models, such as time series models, struggle to accurately capture the nonlinear characteristics of electricity market price fluctuations and the impact of complex factors such as fluctuations in renewable energy output, resulting in low electricity price forecast accuracy.

[0004] In view of this, this application is hereby filed. Summary of the Invention

[0005] The purpose of the present invention is to provide an auxiliary quotation method for the electricity spot day-ahead market to solve the problems of difficult data integration, low prediction accuracy and insufficient risk control mentioned in the above background technology.

[0006] To solve the above technical problems, the present invention provides an auxiliary quotation method for the electricity spot day-ahead market, comprising the following steps: S1 Multi-source Data Fusion Preprocessing: This collects grid operation data, meteorological data, market data, and carbon market data, fills in missing data values ​​using a spatiotemporal generative adversarial network, synchronizes meteorological data in 15-minute cycles, and uses a natural language processing model to extract strategic keywords from market application texts to form a spatiotemporally aligned multidimensional dataset. S2 Grid Topology Constraint Modeling: Constructs a grid topology graph with node attributes and edge weights, embeds grid power balance constraints through a graph neural network, and outputs node electricity price predictions that satisfy physical laws. S3 spatiotemporal feature engineering and decision optimization: This approach uses a time-series neural network combined with an attention mechanism to capture the temporal fluctuations in electricity prices and the spatial correlation between nodes. It also considers the effects of transmission delays to generate a multidimensional feature matrix. Based on a two-layer optimization model, it calculates the optimal futures contract ratio and unit combination plan, and iteratively outputs the final quotation strategy. S4 Carbon-Electricity Joint Optimization: Real-time monitoring of carbon prices. When the preset threshold is exceeded, the quotation amount of the thermal power unit is adjusted during the high and low carbon emission periods, and carbon quota trading recommendations are generated simultaneously. Through the deep integration of multi-source data, grid topology constraint modeling and cross-market collaborative optimization, the problems of traditional quotation methods such as data isolation, inaccurate predictions, and insufficient risk control are solved, effectively improving the accuracy of quotations and market competitiveness, and realizing the maximization of profits and controllable risks of power generation companies.

[0007] Furthermore, in S1, the collected grid operation data includes SCADA real-time bus load, transmission line capacity and unit operation status, market data includes historical day-ahead electricity prices and competitor declaration texts, and carbon market data includes regional carbon quota prices and corporate carbon asset holdings; ensuring the comprehensiveness and diversity of data sources to provide a basis for subsequent accurate analysis.

[0008] Furthermore, in S1, the spatiotemporal generative adversarial network generates missing power grid operation data based on historical data of adjacent sites, and the natural language processing model uses the BERT model to perform entity recognition on the declaration text and extract strategic keywords such as "conservative quotation" and "aggressive volume grabbing"; improve data integrity and availability, enhance insight into competitor strategies, and optimize quotation strategy formulation.

[0009] Furthermore, in S2, the node attributes of the power grid topology diagram include the bus load forecast value and the unit type, and the edge weight is set to the transmission line capacity limit; the power balance constraint is used to force the model to output the electricity price forecast result that conforms to the physical laws of the power grid; the electricity price forecast is made more in line with the actual power grid operation situation, the forecast accuracy is improved, and the quotation deviation is reduced.

[0010] Furthermore, in S3, the futures layer of the two-layer optimization model calculates the optimal futures contract ratio based on option pricing, and the spot layer generates a unit combination plan through mixed integer programming. The two-layer model performs at least three rounds of iterations to achieve coordinated optimization of risks and returns; realizes the linkage between the spot and futures markets, effectively hedges price risks, and improves the comprehensive benefits of power generation enterprises.

[0011] Furthermore, in the S3, the time series neural network adopts the Transformer-LSTM architecture to capture the time series characteristics of electricity prices, and the attention mechanism calculates the spatial association weights of nodes and introduces a 1-2 hour lag time window; it accurately captures the temporal and spatial variation patterns of electricity prices and provides a more reliable basis for quotation strategies.

[0012] Furthermore, in S4, the preset threshold of carbon price is 100 yuan / ton. By adjusting the quotation amount of thermal power units during high carbon emission periods such as the start-up and shutdown stages, priority is given to declaring low-carbon periods such as full-load operation; carbon-electricity coordinated optimization is achieved, carbon trading costs are reduced, and the economic benefits of enterprises in the carbon market environment are improved.

[0013] Furthermore, it also includes S5 dynamic position adjustment: using deep reinforcement learning to monitor spot price fluctuations in real time, triggering position adjustment actions when the price deviation exceeds 10%, and dynamically adjusting the futures position ratio by 5%-20%; timely adjusting strategies when market prices fluctuate to reduce potential losses and enhance the flexibility and adaptability of quotation strategies.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. Through the deep integration of multi-source data and intelligent modeling, a comprehensive technology system covering data processing, electricity price forecasting, and decision optimization has been established. A spatiotemporal generative adversarial network is used to fill missing values ​​in power grid data and synchronize multi-source data. Combined with the BERT model, market strategy keywords are extracted to form a multidimensional dataset aligned in time and space. This overcomes the data silo problem of traditional methods and lays a reliable data foundation for pricing decisions. Furthermore, a graph neural network is used to embed power balance constraints in the power grid. The Transformer-LSTM and attention mechanisms are combined to capture the spatiotemporal characteristics of electricity prices, accurately reflecting the inherent relationship between power grid operation and price fluctuations.

[0015] 2. In terms of decision optimization and risk control, this invention innovatively constructs a two-tiered optimization model to achieve synergy between the spot and futures markets, coupled with deep reinforcement learning to dynamically adjust positions, simultaneously achieving the "high return, low volatility" pricing objectives. Furthermore, a carbon-electricity combined optimization module automatically adjusts the thermal power quotation period when carbon prices exceed a threshold. Combined with carbon quota trading recommendations, this effectively addresses new market demands under the "dual carbon" goals.

[0016] 3. This method achieves cross-platform migration through Docker containerized deployment, meeting the timeliness requirements of day-ahead market quotes. This significantly improves the quote accuracy and market competitiveness of power generation companies, providing power market players with an auxiliary quote solution that balances economic efficiency and risk controllability. This has important practical value in promoting the coordinated development of the power spot market and the carbon market. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a principle block diagram of an auxiliary quotation method for the electricity spot day-ahead market. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] See also Figure 1The present invention provides a technical solution: an auxiliary quotation method for the electricity spot day-ahead market. The auxiliary quotation method for the electricity spot day-ahead market given in this solution uses technologies such as multi-source data fusion, graph neural network modeling, spatiotemporal feature engineering, and two-layer optimization decision-making to achieve accurate quotation, control risks, and increase the profits of market entities.

[0020] Step 1: Multi-source data collection and preliminary processing 1. Collect multi-dimensional raw data, mainly including: Grid operation data: Real-time bus load data (recorded every 15 minutes), transmission line capacity (such as parameters of the IEEE118 node system), and unit operating status (output limit, start-up and shutdown costs) are obtained from the SCADA system.

[0021] Meteorological data: wind speed and light intensity data in numerical weather forecasts, recorded once an hour, with a resolution of 1km×1km.

[0022] Market data: historical day-ahead electricity prices for the past three years (recorded every 15 minutes), and competitor bids (usually in PDF format describing their bidding strategies).

[0023] Carbon market data: regional carbon quota prices (daily data) and the number of carbon assets held by enterprises.

[0024] 2. The necessity of data cleaning and synchronization: If SCADA data is missing, we use a spatiotemporal generative adversarial network (ST-GAN) to fill it in. For example, if wind power output data is missing, we generate appropriate fill-in values ​​based on the wind speeds and historical output curves of three adjacent sites.

[0025] Interpolate the meteorological data from one record per hour to one record per 15 minutes to make it consistent with the electricity price data in time.

[0026] Use the BERT model to perform entity recognition on market declaration texts and find keywords in competitors' quotation strategies, such as "conservative quotation" and "aggressive volume grabbing".

[0027] For example, when collecting data for a power generation company in Guangdong on XX / XX / 20XX, the SCADA system recorded 288 data points (one every 15 minutes) for the Pearl River Basin busbar load. However, 12 of these data points were lost due to communication failures. The NWP data showed an average wind speed of 6 m / s and a sunlight intensity of 400 W / m² on that day. 2Competitor A's application mentioned "prioritizing basic electricity consumption," while Competitor B mentioned "targeting periods with high electricity prices." Using ST-GAN to generate 12 missing output data points, and then interpolating them, a dataset of 288 aligned data points across all time periods was formed.

[0028] It's important to note that power market data is highly spatiotemporally heterogeneous (grid data is synchronized on a 15-minute cycle, while meteorological data is synchronized on a 1-hour cycle). Traditional interpolation methods simply cannot capture the nonlinear relationships within the data. ST-GAN uses a generative adversarial mechanism to learn data distribution, improving data integrity and providing accurate input data for subsequent modeling.

[0029] Step 2: Build a neural network model (GNN) for the power grid topology 1. First set up the power grid topology diagram: The busbars of the IEEE118-node system are considered as nodes in the graph (a total of 118 nodes), and the transmission lines are weighted edges. The weight is the line transmission capacity limit, in MW.

[0030] Add attributes to the node: bus load forecast value (data processed in step 1), unit type (thermal power / wind power / photovoltaic, the marginal cost of different types of units is different).

[0031] 2. To make the model more consistent with actual grid operation, DC power flow equation constraints are added to the GNN message passing mechanism, ensuring that the node electricity prices output by the model must meet the power balance condition: ; Here is the line susceptance, is the node phase angle, is the unit injection power, It's the load.

[0032] 3. Training the GNN model: The input is the spatiotemporal alignment data generated in step 1 (node ​​attributes plus edge weights).

[0033] The output is the electricity price forecast for each node every 15 minutes for the next 24 hours (a total of 96 time points).

[0034] The loss function used is the mean square error (MSE) plus the power flow equation constraint error (weight 0.3).

[0035] For example, in the case of interprovincial power transmission between Guangdong and Guangxi, the grid topology contains 50 key nodes, 30 in Guangdong and 20 in Guangxi. Edge weights are determined based on the actual transmission line capacity; for example, the capacity of the Guangdong-Guangxi DC line is 500 MW. During training, the model predicted an electricity price of 0.6 yuan / kWh at Guangdong node 1. However, verification of the power flow equation revealed that the power injection at Guangxi node 20 did not meet the equilibrium condition. The model automatically adjusted the predicted value to 0.62 yuan / kWh, reducing the power deviation rate to less than 5%. After 100 rounds of training, the electricity price prediction error for this region was significantly reduced compared to traditional methods.

[0036] It's important to note that the existing public document (CN119358180A) only uses GNN for power flow calculations and doesn't integrate it with electricity price forecasting. This step in our solution embeds grid operating rules into the model through a physical constraint layer, addressing the issue of missing spatial features, such as the impact of transmission congestion on node electricity prices. This improves forecasting accuracy compared to unconstrained GNN models.

[0037] Step 3: Spatiotemporal Attention Feature Engineering 1. Modeling in the time dimension: Use the Transformer encoder to process the electricity price series at 96 time points to capture long-term dependencies, such as the price difference between weekdays and weekends.

[0038] Combined with bidirectional LSTM to process short-term price mutations, 15-minute price fluctuation characteristics can be extracted, and price spikes caused by midday load peaks can be found.

[0039] 2. Necessity of modeling spatial dimensions: The attention mechanism calculates the spatial correlation weights between nodes and identifies key influencing nodes. For example, if the influence weight of Guangdong node 1 on Guangxi node 20 exceeds 0.1, the two nodes are considered to have a strong correlation.

[0040] A lag time window is introduced, and the maximum lag is set to 4 time points, or 1 hour, to capture the delay effect of cross-provincial power transmission. For example, changes in hydropower output in Guangxi may take 1 hour to affect electricity prices in Guangdong.

[0041] 3. During feature fusion, the node electricity price prediction value (spatial feature) output by the GNN and the spatiotemporal features extracted by the Transformer-LSTM are combined to form a multidimensional feature matrix containing 118 nodes × 96 time points.

[0042] For example, on a certain day in the summer of 20XX, a GNN predicted that hydropower output at Guangxi node 20 would peak at 2:00 PM. Using the spatiotemporal attention mechanism, it was found that this output change affected the electricity price at Guangdong node 1 with a lag of two time points (30 minutes). The model automatically adjusted the predicted price at Guangdong node 1 at 2:30 PM from 0.58 yuan / kWh to 0.63 yuan / kWh. This resulted in a 1.2% error from the actual market-clearing price, compared to an 8.5% error for a traditional model that didn't account for this lag effect.

[0043] This demonstrates the strong spatiotemporal coupling of electricity prices. Cross-regional power transmission requires physical transmission time. Traditional time series models (such as ARIMA) only process data from a single node, ignoring spatial correlations. This step in our solution dynamically assigns node influence weights through a spatiotemporal attention mechanism, reducing forecast error for the Guangdong-Guangxi cross-provincial market by 15%, demonstrating the necessity of spatial feature modeling.

[0044] Step 4: Multimodal data fusion and decision graph generation 1. When processing unstructured data, use the BERT model to perform sentiment analysis on competitor submissions and generate a strategy-oriented vector (dimension 100). For example, if a competitor is aggressively targeting volume, the corresponding vector's "price sensitivity" dimension will have a value greater than 0.8.

[0045] 2. Constructing a “volume-price benchmark chart”: Using unit output (x-axis, unit MW) and electricity price (y-axis, unit yuan / kWh) as coordinates, the historical winning bid data are marked as colored scatter points (green represents profit points and red represents loss points).

[0046] The DBSCAN algorithm is used to divide the bidding strategy areas into conservative areas (low electricity prices - low risks), aggressive areas (high electricity prices - high risks), and balanced areas (medium electricity prices - medium returns).

[0047] 3. When integrating decision features, the GNN spatiotemporal features (96×118), BERT strategy vector (1×100), and quantity, price, and area labels (1×3) are combined as input features and input into a multi-layer perceptron (MLP) to generate an initial quotation plan.

[0048] Specifically, for a thermal power unit (maximum output 500MW), the electricity price forecast for the next day at 2:00 PM, obtained in step 3, is 0.65 yuan / kWh, which falls within the aggressive zone. BERT analysis also shows that competitor A tends to bid conservatively during this period (strategy vector "price sensitivity" 0.3), while competitor B tends to bid aggressively (0.7). The model automatically adjusts its bid for this period to 0.63 yuan / kWh (slightly lower than the forecast), placing it within the equilibrium zone. This avoids direct competition with competitor B while also outperforming competitor A's conservative bid. Ultimately, the winning bid for this period is 300MW, 50MW more than the traditional bid.

[0049] Of course, traditional methods only use structured data (such as historical electricity prices) and fail to leverage the semantic information in market submissions. This step in our technical solution uses BERT to extract competitor strategies and then integrates them with a volume-price benchmark chart for visual decision-making. This improves the market adaptability of bidding solutions and addresses the strategic blindness caused by "data silos."

[0050] Step 5: Solve the two-level optimization decision model Upper layer: Futures market risk hedging layer 1. To calculate the optimal futures contract ratio, use the Black-Scholes option pricing model and input these parameters: Spot price volatility σ = 20% (calculated from the historical data in step 3).

[0051] The risk-free interest rate r=3% (yield).

[0052] The futures contract expires in 1 day.

[0053] The objective function is to maximize the expected return minus 1.5 times the conditional value at risk (CVaR), balancing return and risk: ; Here is the futures contract volume, is the futures price, is the initial spot price; Lower layer: spot market unit combination layer 2. When generating the unit combination plan, input the initial bid plan from step 4, unit parameters (startup and shutdown cost of 100,000 yuan / time, coal consumption curve), and grid constraints (node ​​capacity limit). Use the mixed integer programming (MIP) model, and the objective function is to maximize the day-ahead revenue: ; (in yes Time period output, It's a quote. is the cost of electricity generation, is the start-stop state); 3. During two-layer iterative optimization, the optimal futures ratio q calculated by the upper layer is fed back to the lower layer to adjust the risk tolerance of the spot price quote. The unit combination results generated by the lower layer (such as the start-stop plan) are then used to correct the CVaR calculation of the upper layer. Convergence is achieved after three rounds of iteration.

[0054] For example, a power retail company in Zhejiang Province used this process in January 20XX: The upper layer calculated the optimal futures contract ratio to be 40% of the spot trading volume (corresponding to a futures price of 0.55 yuan / kWh and a predicted average spot price of 0.6 yuan / kWh). The lower layer's MIP model adjusted its bidding strategy based on this ratio, increasing output during periods of high electricity prices (forecasted > 0.65 yuan / kWh) and shutting down during periods of low prices (< 0.5 yuan / kWh). As a result, the company's day-ahead bid volume increased from 65% to 82%, resulting in a 0.03 yuan increase in revenue per kilowatt-hour compared to a pure spot trading strategy, and reduced its return volatility from 25% to 12%.

[0055] Clearly, the existing public document (CN202411554063) only optimizes the spot market and doesn't consider futures hedging. This solution utilizes a two-tier model to achieve cross-market collaboration. Monte Carlo simulations have shown that, in scenarios with price fluctuations, such as ±30%, overall returns are increased and risks are reduced compared to single-tier optimization, demonstrating the creative value of cross-market collaboration.

[0056] Step 6: Dynamic Adjustment and Execution Output 1. Real-time market monitoring is crucial. Power grid SCADA data is received in real time through Kafka (with a latency of less than 500ms) and electricity price forecasts are updated every 15 minutes (using the model from steps 2 to 4 for rolling forecasts).

[0057] 2. Use a deep Q-network (DQN) to dynamically adjust spot and futures positions. The state space includes the current position, the deviation of the real-time electricity price (current price / forecast price - 1), and the price fluctuation of carbon quotas. The action space includes five actions, such as increasing the spot position by 10% and reducing the futures position by 10%. The reward function is the immediate return plus 0.9 times the expected future return.

[0058] 3. Regarding carbon-electricity combined optimization, if the real-time carbon price is greater than 100 yuan / ton and the unit is thermal power, the system automatically reduces bids during high-carbon emission periods (such as startup and shutdown), prioritizing bids during low-carbon periods (full-load operation). The final bid output includes unit output quotes every 15 minutes for 24 hours, futures position ratios, and carbon quota trading recommendations.

[0059] For example, at 10:00 AM on XX / XX / 20XX, the Guangdong spot price updated in real time to 0.7 yuan / kWh, higher than the predicted value of 0.65 yuan / kWh in step 5. The DQN model triggered a "reduce futures position" action, lowering the futures ratio from 40% to 30%, freeing up 10% of electricity for spot trading. At the same time, because the current carbon price was 120 yuan / ton, the model recommended that the thermal power company increase its bid by 0.02 yuan / kWh during the full-load operation period from 2:00 PM to 4:00 PM (leveraging its low-carbon advantages to capture the green electricity premium). The winning bid for this period was 0.72 yuan / kWh, resulting in an additional profit of 120,000 yuan compared to the original plan, and a net increase of 50,000 yuan in carbon allowance revenue.

[0060] Of course, market price fluctuations and carbon policy changes are so rapid that traditional static pricing simply cannot cope. This step of our solution, through real-time DQN strategy adjustment, reduced losses by 40% during a pilot program in a city's electricity spot market when experiencing extreme price fluctuations (>0.8 yuan / kWh). Combined with carbon-electricity joint optimization, this reduced profit volatility for integrated energy service providers by 30%-40%, demonstrating the practical benefits of our dynamic adjustments.

[0061] In summary, in terms of data processing and modeling, the combination of a spatiotemporal generative adversarial network (ST-GAN) and the BERT natural language processing model addresses the heterogeneity and integration challenges of multi-source data in the power market. This creates a multidimensional dataset aligned in time and space, providing comprehensive data support for bidding decisions. Furthermore, the first-ever integration of a graph neural network (GNN) with power balance constraints in the power grid allows for the construction of a topological model that accounts for the physical laws of the power grid. This results in lower electricity price forecast errors than traditional methods, overcoming the bottleneck of traditional forecasting models that cannot accurately reflect the actual operating status of the power grid.

[0062] In terms of decision-making optimization mechanisms, an innovative two-layer optimization model was constructed, comprising a futures risk hedging layer and a spot unit portfolio layer. This enabled collaborative decision-making between the electricity spot and futures markets. By combining option pricing with mixed integer programming, the system increased the proportion of electricity awarded to power generation companies and reduced return volatility, effectively addressing the inadequate risk control inherent in traditional bidding methods. Furthermore, the system introduced deep reinforcement learning to dynamically adjust positions, triggering strategy corrections when market price fluctuations exceed 10%. This reduced losses in extreme scenarios and significantly enhanced the market adaptability of the bidding strategy.

[0063] In response to the new demands of the "dual carbon" goals, a carbon-electricity joint optimization module is proposed. When the carbon price exceeds 100 yuan / ton, the quotation amount of thermal power units during high and low carbon emission periods is automatically adjusted. Combined with carbon quota trading recommendations, the comprehensive profit per kilowatt-hour is increased by 3%-5%, realizing the coordinated optimization of electricity market quotations and carbon markets, and providing a new solution for the quotation decisions of power generation companies under the "dual carbon" goals.

Claims

1. A method for auxiliary quotation in the electricity spot day-ahead market, characterized by: The following steps are involved: S1 Multi-source Data Fusion Preprocessing: This collects grid operation data, meteorological data, market data, and carbon market data, fills in missing data values ​​using a spatiotemporal generative adversarial network, synchronizes meteorological data in 15-minute cycles, and uses a natural language processing model to extract strategic keywords from market application texts to form a spatiotemporally aligned multidimensional dataset. S2 Grid Topology Constraint Modeling: Constructs a grid topology graph with node attributes and edge weights, embeds grid power balance constraints through a graph neural network, and outputs node electricity price predictions that satisfy physical laws. S3 spatiotemporal feature engineering and decision optimization: This approach uses a time-series neural network combined with an attention mechanism to capture the temporal fluctuations in electricity prices and the spatial correlation between nodes. It also considers the effects of transmission delays to generate a multidimensional feature matrix. Based on a two-layer optimization model, it calculates the optimal futures contract ratio and unit combination plan, and iteratively outputs the final quotation strategy. S4 Carbon-Electricity Joint Optimization: Real-time monitoring of carbon prices. When the preset threshold is exceeded, the quoted amount of the thermal power unit during high and low carbon emission periods is adjusted, and carbon quota trading recommendations are generated simultaneously.

2. The auxiliary quotation method for the electricity spot day-ahead market according to claim 1, characterized in that: In S1, the collected grid operation data include SCADA real-time bus load, transmission line capacity and unit operation status, market data include historical day-ahead electricity prices and competitor declaration texts, and carbon market data include regional carbon quota prices and corporate carbon asset holdings.

3. The auxiliary quotation method for the electricity spot day-ahead market according to claim 1, characterized in that: In S1, the spatiotemporal generative adversarial network generates missing power grid operation data based on the historical data of adjacent sites, and the natural language processing model uses the BERT model to perform entity recognition on the declaration text and extract the keywords of the "conservative quotation" and "aggressive volume grabbing" strategies.

4. The auxiliary quotation method for the electricity spot day-ahead market according to claim 1, characterized in that: In S2, the node attributes of the power grid topology include the bus load forecast value and the unit type, and the edge weight is set to the transmission line capacity limit; the power balance constraint is used to force the model to output the electricity price forecast result that conforms to the physical laws of the power grid.

5. The auxiliary quotation method for the electricity spot day-ahead market according to claim 1, characterized in that: In S3, the futures layer of the two-layer optimization model calculates the optimal futures contract ratio based on option pricing, and the spot layer generates a unit commitment plan through mixed integer programming.

6. The auxiliary quotation method for the electricity spot day-ahead market according to claim 1, characterized in that: In S3, the time series neural network uses the Transformer-LSTM architecture to capture the time series characteristics of electricity prices, and the attention mechanism calculates the spatial association weights of nodes and introduces a 1-2 hour lag time window.

7. The auxiliary quotation method for the electricity spot day-ahead market according to claim 1, characterized in that: In S4, the preset threshold of carbon price is RMB 100 / ton. By adjusting the quotation amount of the thermal power units during the high carbon emission period during the start-up and shutdown stages, priority is given to declaring the low-carbon period of full-load operation.

8. The auxiliary quotation method for the electricity spot day-ahead market according to claim 1, characterized in that: It also includes S5 dynamic position adjustment: using deep reinforcement learning to monitor spot price fluctuations in real time, triggering position adjustment actions when the price deviation exceeds 10%.

Images