Medium and long term price prediction method and device, electronic equipment and storage medium

Abstract

The application relates to the field of power grid dispatching, and provides a medium and long term price prediction method and device, an electronic device and a storage medium. The method comprises the following steps: taking a to-be-predicted date as a starting time point, and collecting historical multi-dimensional feature data sets in reverse; determining power boundary feature data corresponding to the to-be-predicted date based on the historical multi-dimensional feature data sets; screening a target training data set from the historical multi-dimensional feature data sets based on the power boundary feature data and the historical multi-dimensional feature data sets; training an initial medium and long term price prediction model by using the target training data set, so as to obtain a trained medium and long term price prediction model; and inputting multi-dimensional feature data corresponding to the to-be-predicted date into the trained medium and long term price prediction model, and outputting a medium and long term price prediction result. The application effectively solves the problems that the existing technology has poor ability to describe complex market dynamics, the prediction accuracy is difficult to break through the existing bottleneck, and it is difficult to realize high-precision and high-reliability prediction of medium and long term prices.

Description

Technical Field

[0001] This application relates to the field of power grid dispatching, and in particular to a medium- and long-term price forecasting method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the booming development of the electricity spot market, accurate prediction of medium- and long-term electricity prices has become a core and crucial element for all participants in the electricity market to enhance their market competitiveness, conduct marketing, optimize trading bidding strategies, ensure electricity revenue, and reduce energy costs.

[0003] Traditional medium- and long-term price forecasting methods are mostly based on statistical models and machine learning models, such as time series models (ARIMA) and support vector machine models (SVM). These methods often only use historical electricity price data and fail to effectively integrate multi-source heterogeneous data that affect electricity prices (such as meteorological data, renewable energy output data, grid congestion information, market policy documents, etc.), which leads to insufficient ability to characterize complex market dynamics and difficulty in breaking through existing bottlenecks in forecast accuracy.

[0004] Therefore, improving the accuracy of depicting complex market dynamics, breaking through the existing bottlenecks in forecasting precision, and achieving high-precision and high-reliability forecasting of medium- and long-term prices to provide technical support for electricity market transactions is one of the urgent technical challenges to be solved in this field. Summary of the Invention

[0005] In view of this, embodiments of this application provide a medium- to long-term price forecasting method, apparatus, electronic device, and storage medium to address the problems in the prior art where the ability to characterize complex market dynamics is insufficient, the forecasting accuracy is difficult to break through existing bottlenecks, and it is difficult to achieve high-precision and high-reliability forecasting of medium- to long-term prices, thus providing technical support for electricity market transactions.

[0006] A first aspect of this application provides a medium- to long-term price forecasting method, comprising: Starting from the date to be predicted, historical multi-dimensional feature datasets are collected backward. The historical multi-dimensional feature dataset includes multiple historical multi-dimensional feature data, and each historical multi-dimensional feature data includes power system operation feature data, meteorological environment feature data, and power market transaction feature data for its corresponding historical date. Based on historical multi-dimensional feature datasets, determine the power boundary feature data corresponding to the date to be predicted; Based on power boundary feature data and historical multi-dimensional feature datasets, a target training dataset is selected from the historical multi-dimensional feature datasets. The initial medium- and long-term price prediction model is trained using the target training dataset until the preset convergence condition is met, resulting in a well-trained medium- and long-term price prediction model. Input the multi-dimensional feature data corresponding to the date to be predicted into the trained medium- and long-term price prediction model, and output the medium- and long-term price prediction results.

[0007] A second aspect of this application provides a medium- to long-term price forecasting device, comprising: The data acquisition module is configured to collect historical multi-dimensional feature datasets by tracing back from the date to be predicted. The historical multi-dimensional feature dataset includes multiple historical multi-dimensional feature data, and each historical multi-dimensional feature data includes power system operation feature data, meteorological environment feature data, and power market transaction feature data for its corresponding historical date. The determination module is configured to determine the power boundary feature data corresponding to the date to be predicted based on a historical multi-dimensional feature dataset. The filtering module is configured to filter out the target training dataset from the historical multi-dimensional feature dataset based on power boundary feature data and historical multi-dimensional feature dataset. The training module is configured to train an initial medium- to long-term price prediction model using the target training dataset until a preset convergence condition is met, resulting in a well-trained medium- to long-term price prediction model. The prediction module is configured to input multi-dimensional feature data corresponding to the date to be predicted into a trained medium- to long-term price prediction model and output medium- to long-term price prediction results.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0010] Compared with existing technologies, the beneficial effects of this application's embodiments include at least the following: It innovatively proposes a cascaded prediction framework of "predicting boundaries first, then predicting prices," specifically divided into two interconnected stages. Each stage performs its own function and progresses step-by-step to achieve accurate prediction of medium- and long-term electricity prices. The specific details are as follows: The first stage is the electricity boundary feature prediction stage: Starting from the date to be predicted, historical multi-dimensional feature datasets are collected backward. Based on this historical multi-dimensional feature dataset, the electricity boundary feature data corresponding to the date to be predicted is predicted. This stage effectively solves the technical pain point in existing technologies for medium- and long-term price prediction, where the inability to obtain key boundary information such as grid load and renewable energy output on the date to be predicted in advance leads to a lack of core input features in the prediction model, thus making it difficult to guarantee prediction accuracy. The second stage is the accurate electricity price prediction stage: Based on the electricity boundary feature prediction data output from the first stage, combined with the aforementioned historical multi-dimensional feature dataset, a target training dataset is selected from the historical multi-dimensional feature dataset. This screening method achieves high-fidelity alignment between the target training dataset and the future prediction scenario, which not only greatly improves the scenario adaptability and prediction accuracy of the prediction model, but also enables the rules learned by the model to be directly adapted to specific future prediction scenarios, significantly enhancing the relevance, accuracy and reliability of the prediction results.

[0011] Through the aforementioned two-stage cascaded design, the complex task of electricity price forecasting is decoupled into two relatively independent and easier-to-model sub-tasks: boundary forecasting and price forecasting. This avoids the problems of extremely complex model structures and overfitting inherent in existing technologies that construct a single "black box" model (i.e., directly mapping long-term meteorological characteristics to electricity prices). Furthermore, this cascaded framework has a clear structure, significantly improving the model's overall generalization ability, operational stability, and interpretability of prediction results, effectively overcoming the limitations of existing forecasting methods. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a medium- to long-term price forecasting method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a medium- to long-term price forecasting device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0015] A medium- to long-term price forecasting method and apparatus according to embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart illustrating a medium- to long-term price forecasting method provided in an embodiment of this application. Figure 1 Medium- to long-term price forecasting methods can be executed by a server. For example... Figure 1 As shown, this medium- to long-term price forecasting method includes the following steps: Step S101: Starting from the date to be predicted, collect historical multi-dimensional feature datasets by tracing back. The historical multi-dimensional feature datasets include multiple historical multi-dimensional feature data. Each historical multi-dimensional feature data includes power system operation feature data, meteorological environment feature data, and power market transaction feature data for its corresponding historical date. Step S102: Based on the historical multi-dimensional feature dataset, determine the power boundary feature data corresponding to the date to be predicted; Step S103: Based on the power boundary feature data and the historical multi-dimensional feature dataset, select the target training dataset from the historical multi-dimensional feature dataset; Step S104: Train the initial medium- and long-term price prediction model using the target training dataset until the preset convergence condition is met, and obtain the trained medium- and long-term price prediction model. Step S105: Input the multi-dimensional feature data corresponding to the date to be predicted into the trained medium- and long-term price prediction model, and output the medium- and long-term price prediction results.

[0017] Power system operation characteristic data includes day-ahead boundary characteristics on both the supply and demand sides, such as central dispatch load, tie line load, and renewable energy output, as well as grid operation status information such as node congestion and fault maintenance.

[0018] Meteorological environmental characteristic data, including meteorological parameters (or meteorological elements) such as 100 m wind speed, surface solar radiation downflow, temperature, and humidity.

[0019] Electricity market transaction characteristic data, including market transaction-related data such as the unified clearing price across the entire network.

[0020] In scenarios such as power system dispatching, power market trading, and load or electricity price forecasting, power boundary characteristic data refers to the key boundary parameters and characteristic data that define the operating status, supply and demand balance conditions, and market trading constraints of the power system. It serves as the basis for analyzing system operating trends and formulating trading strategies.

[0021] Electricity boundary characteristic data mainly includes supply and demand boundary information and operational constraint boundary information. Supply and demand boundary information determines the balance between electricity supply and demand in the power system and is a core input feature for load forecasting and electricity price forecasting. Supply and demand boundary information mainly includes: 1) Demand side: load-related data such as central dispatch load, regional load, and user electricity consumption plans; 2) Supply side: power source and transmission-related data such as renewable energy output plans (photovoltaic, wind power), thermal power unit operating capacity, hydropower output ceiling, and tie-line transmission power. Operational constraint boundary information represents the constraints on the safe and stable operation of the power system and indirectly affects market clearing prices and trading strategies. Operational constraint boundary information mainly includes: 1) Grid constraints: node congestion status, transmission line capacity limits, transformer operating thresholds, etc.; 2) Maintenance constraints: generator unit maintenance plans, transmission line maintenance arrangements, etc.; 3) Market rule constraints: policy-related boundary conditions such as the upper limit of electricity spot or medium- and long-term trading prices, transaction rules, and settlement standards.

[0022] The preset convergence condition can be that the number of iterations reaches a preset threshold (which can be set according to the actual situation, such as 50 rounds, 100 rounds, 200 rounds, etc.), or the target metric of the validation set (such as RMSE) does not decrease in K consecutive iterations.

[0023] In the training process of medium- and long-term price forecasting models, the core purpose of setting convergence conditions is to balance model accuracy and training efficiency, and to avoid overfitting or ineffective iterations. In practical applications, appropriate convergence conditions can be set according to the model type and business needs; no specific restrictions are imposed in this application.

[0024] As an example, for the date to be predicted Using the base date (or labeled date), we trace back N days (i.e., N records, where the value of N can be set according to the actual situation, for example, 60, 70, 80, etc.) of historical multi-dimensional feature data, which can be labeled as historical multi-dimensional feature data. , , ... ,in, This indicates the day before the predicted date. This indicates the day before the predicted date. This indicates the 3rd day before the predicted date, and so on. This represents the Nth day before the date to be predicted. For example, the date to be predicted... October 10, 2025 October 9, 2025 October 8, 2025 October 7, 2025.

[0025] The technical solution provided in this application innovatively proposes a cascaded prediction framework of "predicting the boundary first, then predicting the price," comprising two interconnected stages: the first stage is the electricity boundary feature prediction stage, and the second stage is the accurate electricity price prediction stage. Through this two-stage cascaded design, the complex electricity price prediction task is decoupled into two relatively independent and easier-to-model sub-tasks: boundary prediction and price prediction. This avoids the problems of extremely complex model structures and overfitting inherent in existing technologies that construct a single "black box" model (i.e., directly mapping long-term meteorological characteristics to electricity prices). Simultaneously, the clear structure of this cascaded framework significantly improves the overall generalization ability, operational stability, and interpretability of the prediction results, effectively overcoming the limitations of existing prediction methods.

[0026] In some embodiments, based on a historical multi-dimensional feature dataset, the power boundary feature data corresponding to the date to be predicted is determined, including: Determine the historical medium- to long-term electricity price corresponding to each historical date; Target multi-dimensional feature data is obtained by filtering historical multi-dimensional feature data based on historical medium- and long-term electricity prices. Based on the target's multi-dimensional feature data, determine the power boundary feature data corresponding to the date to be predicted.

[0027] Historical medium- and long-term electricity prices refer to the collection of medium- and long-term electricity trading prices that have been actually traded, settled, or cleared in the electricity market before the forecast date (base date), including but not limited to fixed historical transaction price data such as medium- and long-term contract electricity prices, monthly or annual bilateral negotiated prices, centralized trading prices, and listed trading prices.

[0028] In some implementations, feature engineering is performed on each historical multi-dimensional feature data in the historical multi-dimensional feature dataset to obtain the corresponding historical power boundary feature data, specifically including: For each historical multi-dimensional feature data, feature filtering and reconstruction are performed on the day-ahead boundary features on both the supply and demand sides to obtain combined features that are strongly correlated with medium- and long-term price forecasts. For power grid operation status information, the occurrence of an event is used as the judgment condition. A quantitative feature is constructed through one-hot encoding to obtain the power grid operation status encoding feature matrix. Feature screening and reconstruction of meteorological environmental characteristic data yielded meteorological characteristics of the entire province that are strongly correlated with medium- and long-term price forecasts. Construct time-lag instantaneous features and period-lag mean features based on electricity market transaction characteristic data; The timestamps of historical multi-dimensional feature data are parsed and quantized to obtain a time feature set. By combining the combined features, the power grid operation status coding feature matrix, the provincial meteorological features, the point-in-time lagging instantaneous features, the periodic lagging mean features, and the time feature set, the historical power boundary feature data corresponding to the historical date of this historical multi-dimensional feature data is obtained.

[0029] In some implementations, the day-ahead boundary features on both the supply and demand sides are screened and reconstructed to obtain combined features that are strongly correlated with medium- and long-term price forecasts, specifically including: The day-ahead boundary characteristics on both the supply and demand sides are classified to obtain the day-ahead boundary characteristics on the demand side and the day-ahead boundary characteristics on the supply side. By linearly combining the day-ahead boundary characteristics of the demand side and the day-ahead boundary characteristics of the supply side, a combined characteristic strongly correlated with medium- and long-term price forecasts is obtained.

[0030] As an example, the day-ahead boundary features on both the supply and demand sides are classified according to business attribute labels (such as demand or supply), resulting in demand-side and supply-side day-ahead boundary features. Demand-side features are marked as "positive," and their original values ​​(or weighted original values) are directly retained when participating in linear combinations. Supply-side features are marked as "negative," and their representation is shown as a "subtraction term" when participating in linear combinations. Then, the demand-side and supply-side features are normalized to obtain normalized values. A first non-negative weight is assigned to the normalized demand-side features, and a second non-negative weight is assigned to the normalized supply-side features. Finally, a linear combination is performed based on the normalized demand-side features, the first non-negative weight, the normalized supply-side features, and the second non-negative weight to obtain a combined feature strongly correlated with medium- to long-term price forecasts.

[0031] In some implementations, the power grid operating status information is quantized using the occurrence of an event as a criterion, and a power grid operating status coding feature matrix is ​​obtained by constructing quantization features through one-hot coding. Specifically, this includes: First, for the power grid operation status information to be quantified, independent classification dimensions are decomposed according to the binary judgment logic of "whether it has occurred," and each type of status is assigned a unique identifier. For example, assume the power grid operation status information to be quantified includes five types of information: ① transmission line congestion events; ② generator unit fault maintenance events; ③ substation equipment abnormality events; ④ power flow exceeding limits events; ⑤ tie line power exceeding limits events. Each of these types of information is treated as an independent classification dimension, and the dimension identifiers are denoted as A1, A2, A3, A4, and A5, respectively. Then, a binary label matrix of power grid operation status is constructed. Specifically, for each type of power grid operation status at each time sample, a binary assignment is performed according to "whether it has occurred": if the corresponding status at that time point actually occurs, the label value is assigned as 1; if it has not occurred, the label value is assigned as 0. For example, if a transmission line blockage occurs on March 1, 2024 (A1=1), no generator maintenance occurs (A2=0), substation equipment is normal (A3=0), power flow does not exceed limits (A4=0), and tie line power is normal (A5=0), then the label vector corresponding to this time row is [1,0,0,0,0]. Next, based on the above binary label matrix, one-hot encoding is performed on each column (power grid operating state dimension) to ensure that non-numerical state information is converted into quantifiable features that the model can recognize. This generates a power grid operating state encoding feature matrix with dimensions of "number of time samples × number of power grid operating state dimensions," where all elements are either 0 or 1, thus completing the quantification of power grid operating state information.

[0032] In some implementations, meteorological environmental characteristic data are subjected to feature screening and reconstruction to obtain provincial meteorological characteristics that are strongly correlated with medium- and long-term price forecasts, specifically including: Meteorological environmental characteristic data are divided into meteorological characteristics strongly correlated with new energy sources and general meteorological characteristics; For meteorological features strongly correlated with new energy, the first reconstructed meteorological features are obtained by weighting and reconstructing them according to the grid-based new energy installed capacity ratio of the provinces to be tested. For general meteorological features, area-weighted uniform reconstruction is used to obtain the second reconstructed meteorological features; The first and second reconstructed meteorological features were subjected to feature scale transformation and normalization to obtain the provincial meteorological features that are strongly correlated with medium- and long-term prices.

[0033] Meteorological characteristics strongly correlated with new energy sources include 100-meter wind speed, downward solar radiation flux at the surface, and wind direction, which directly affect wind and solar power output. General meteorological factors include temperature, humidity, air pressure, and precipitation, which indirectly affect power load and grid operation.

[0034] As an example, for meteorological features strongly correlated with new energy, the first reconstructed meteorological features are obtained by weighting and reconstructing them according to the proportion of new energy installed capacity in the grid of the province to be predicted. Specifically, this includes: calculating the proportion of new energy installed capacity in each grid of the province to be predicted (such as the proportion of wind power installed capacity and the proportion of photovoltaic installed capacity); and weighting and reconstructing the meteorological features strongly correlated with new energy based on the proportion of new energy installed capacity to obtain the first reconstructed meteorological features.

[0035] By performing feature scale transformations (such as linear transformations, nonlinear transformations, or logarithmic transformations) on the first and second reconstructed meteorological features mentioned above, the correlation between meteorological elements and medium- and long-term price forecasts can be improved.

[0036] Normalizing the transformed first and second reconstructed meteorological features can eliminate the dimensional differences between the transformed first and second reconstructed meteorological features and the medium- and long-term price forecasts, enhance the correlation of the data, and obtain provincial meteorological features that are strongly correlated with the medium- and long-term price forecasts.

[0037] In some implementations, time-lag instantaneous features and period-lag mean features are constructed based on electricity market transaction characteristic data, specifically including: Using the daily unified clearing price as the basic variable, and arranged in calendar day order, a daily time series is formed: Where t represents the date to be predicted, This indicates the uniform clearing price for the predicted date. This indicates the uniform clearing price for the day preceding the predicted date. This indicates the uniform clearing price on the second day prior to the predicted date. This represents the unified clearing price N days before the predicted date. Then, using the predicted date as a reference, the original instantaneous price values ​​from day 1 to day 45 before the predicted date are taken sequentially to construct a 45-dimensional time-lag instantaneous feature. Each lag day corresponds to an independent feature, resulting in a total of 45-dimensional time-lag instantaneous features, which are used to capture recent daily price fluctuations and abrupt changes.

[0038] Based on the lagged instantaneous value, the average clearing price within each period is calculated using daily, weekly, and monthly periods, forming a periodic lagged mean characteristic. This periodic lagged mean characteristic includes daily, weekly, and monthly periodic lagged mean characteristics. For example, the value of the lagged mean before the date to be predicted is taken. i The price of the day is used as the daily average (equivalent to the instantaneous value of the day) to construct the daily cyclical average feature. A 7-day periodic window is used, and the price is calculated based on the price before the date to be predicted. i The average of the 7 consecutive days in which the day is located is used to reflect the price center level of the past week. A monthly cycle window of 28-31 days is used, with the day preceding the forecast date...i The average of the 30 consecutive days in which the price is located is used to reflect the price trend and overall level over the past month.

[0039] Finally, the features are aligned and stitched together on the same benchmark date time dimension to form a complete set of lagged features corresponding to a unified clearing price.

[0040] In some implementations, the timestamps of historical multi-dimensional feature data are parsed and quantized to obtain a time feature set, specifically including: The timestamp is parsed to obtain time-dimension, week-dimension, month-dimension, and year-dimension information; Sine and cosine encoding is performed on the time-dimensional information to obtain the time-dimensional encoded features; One-hot encoding is performed on the weekly, monthly, and yearly information respectively to obtain the corresponding weekly, monthly, and yearly encoded features.

[0041] The time dimension information reflects the details of time within a day, and can be data at the 15-minute level or the hour level. If it is data at the 15-minute level, it includes not only the hour and minute, but also the corresponding time point index value (e.g., 0~95, corresponding to 96 15-minute time points in 1 day). If it is data at the hour level, it includes the hour (index value 0~23).

[0042] Weekly dimension information, the attributes related to "week" correspond to "day of the week". Optional extraction of week number (to assist in analyzing weekly trends) aligns with the weekly cycle patterns of medium- to long-term prices.

[0043] Monthly dimension information includes attributes related to "month" and corresponding month numbers. Optional extraction dates and number of days in the month can be selected to align with the monthly cycle patterns of medium- to long-term prices (seasons, trading plans, etc.).

[0044] Year-related information includes attributes associated with the year and corresponding year numbers. Optional extraction of year date sequences can aid in analyzing long-term trends.

[0045] As an example, the timestamp is "2025-01-10 08:15:00". Parsing this timestamp, the time dimension information is 8 hours and 15 minutes, and the corresponding time point index is 33, that is, the 33rd 15-minute point (08:15); the week dimension information is Friday, the second week of 2025; the month dimension information is January, the 10th day of the month; and the year dimension information is 2025, the 10th day of the year.

[0046] The combined features obtained after the above feature engineering, the power grid operation status coding feature matrix, the provincial meteorological features, the point-in-time lagging instantaneous features, the periodic lagging mean features, and the time feature set are combined to obtain the processed historical multidimensional feature data corresponding to this historical multidimensional feature data.

[0047] By constructing a complete feature engineering system that includes instantaneous features with time lag, mean features with periodic lag, a power grid operation status coding feature matrix, provincial meteorological features, and time feature sets, it is possible to simultaneously capture short-term abrupt changes, periodic patterns, long-term trends, and the impact of power grid operation constraints on electricity prices, effectively improving the feature's ability to represent the electricity price formation mechanism. By combining lag features and periodic features, the model's ability to learn time-series patterns is enhanced, avoiding prediction bias caused by insufficient information from a single feature. By quantifying discrete information such as power grid operation status through one-hot coding, unstructured information can be modeled, improving the completeness and standardization of model input. The overall feature engineering is compatible with the cascaded prediction framework, making the model easier to converge, with stronger generalization ability, and significantly improved prediction accuracy and stability.

[0048] Next, the processed historical multi-dimensional feature data is filtered based on historical medium- and long-term electricity prices to obtain feature data strongly correlated with historical medium- and long-term electricity prices, i.e., target multi-dimensional feature data. For example, for each historical date, the correlation coefficient (e.g., Pearson correlation coefficient) between the historical medium- and long-term electricity price corresponding to that historical date and each type of feature in the processed historical multi-dimensional feature data is calculated. Feature data with correlation coefficients greater than a preset threshold (which can be flexibly set according to actual conditions, such as 0.6, 0.7, etc.) are classified as target multi-dimensional feature data. Then, the target multi-dimensional feature data is used as input, and a machine learning model (e.g., XGBoost model) is used for prediction to obtain the electricity boundary feature data corresponding to the date to be predicted.

[0049] By utilizing historical multi-dimensional feature data that can be obtained reliably over a long period, the market supply and demand boundary conditions for the forecast date can be accurately calculated, providing complete and high-quality input features for subsequent price forecasting models. This fundamentally solves the problem of insufficient input data sources in medium- and long-term price forecasting, breaks through the technical bottleneck of missing power boundary feature data for the forecast date in medium- and long-term price forecasting, and realizes a technical leap from difficult to accurately predict medium- and long-term power prices to stable and accurate prediction.

[0050] In some embodiments, based on power boundary feature data and historical multi-dimensional feature datasets, a target training dataset is selected from the historical multi-dimensional feature dataset, including: The target bidding space vector is determined based on power boundary feature data, and the historical bidding space vector set is determined based on historical multi-dimensional feature datasets. Based on the target bidding space vector and the historical bidding space vector set, the target training dataset is selected from the historical multi-dimensional feature dataset.

[0051] In some implementations, the target bidding space vector is determined based on power boundary feature data, including: Select all boundary data reflecting electricity demand, such as centrally dispatched load data and regional electricity load data, and denote them as the demand set {D1, D2, ..., D...}. m}, where m is the number of dimensions of the demand-side boundary data, and each data point represents the power boundary characteristic data corresponding to the date to be predicted. All boundary data reflecting power supply, including renewable energy output data (photovoltaics, wind power, etc.), available output data of thermal power units, hydropower output data, and tie-line input power data, are selected and denoted as the supply set {S1, S2, ..., S}. r}, where r is the number of dimensions of the supply-side boundary data, and the data type is consistent with that of the demand side (actual / predicted data correspondence and match, conforming to the output results of the feature engineering and boundary prediction stages mentioned above). Linear summation is performed on the demand-side set and the supply-side set respectively to obtain the total demand and total supply for the date to be predicted, eliminating the one-sidedness of single-dimensional data and providing a basis for subsequent bidding space calculation. Next, based on the calculation results of total demand and total supply, the linear calculation method of "total demand minus total supply" is adopted to obtain the target bidding space vector corresponding to the date to be predicted. For example, the target bidding space vector is calculated according to formula (1).

[0052] (1); In equation (1), This represents the target bidding space vector for the date to be tested. This represents the total demand for the predicted date. This represents the total supply for the forecast date.

[0053] when When the value is greater than 0, it indicates that the power system is in a state of demand exceeding supply, the target bidding space vector is positive, reflecting the intensity of market bidding; when... When the value is ≤0, it indicates that the power system is in a state where supply is greater than or equal to demand. The target bidding space vector is negative, reflecting the degree of leniency in market bidding. The calculation results retain reasonable precision (such as retaining 1 decimal place) to meet the requirements of constructing the target bidding space vector and calculating similarity in subsequent steps.

[0054] Understandably, referring to the method for determining the target bidding space vector mentioned above, for each historical multi-dimensional feature data, the historical bidding space vector corresponding to each historical date is calculated, and then arranged in chronological order to form a set of historical bidding space vectors.

[0055] In some embodiments, the historical bidding space vector set includes multiple historical bidding space vectors, and one historical bidding space vector corresponds to one historical date; Based on the target bidding space vector and the historical bidding space vector set, a target training dataset is selected from the historical multi-dimensional feature dataset, including: Calculate the first and second similarities between the target bidding space vector and each historical bidding space vector; For each historical date, the comprehensive similarity between the historical date and the date to be predicted is calculated based on the first and second similarities corresponding to the historical date. Based on the comprehensive similarity corresponding to each historical date, the target training dataset is selected from the historical multi-dimensional feature dataset.

[0056] For ease of description, we continue with the above example, calculating the first similarity (e.g., Euclidean distance) and the second similarity (e.g., Pearson correlation coefficient) between the target bidding space vector and each historical bidding space vector in the historical bidding space vector set. Then, we sort the first similarity corresponding to each historical date according to Euclidean distance from smallest to largest, obtaining the first similarity ranking result; and sort the second similarity corresponding to each historical date according to Pearson correlation coefficient from smallest to largest, obtaining the second similarity ranking result; finally, we perform a weighted calculation on the first similarity ranking result and the second similarity corresponding to each historical date to obtain the comprehensive similarity between each historical date and the date to be predicted.

[0057] As an example, suppose the historical multi-dimensional feature dataset includes 60 historical multi-dimensional feature data points from days 1 to 60 before the date to be predicted, in order to calculate the historical date. (The day before the predicted date) and the predicted date Let's take the comprehensive similarity between them as an example for a detailed explanation. Assume historical dates... Corresponding historical bidding space vector and the date to be predicted The first similarity between the corresponding target bidding space vectors is ranked as 1 in the first similarity ranking result (the smaller the Euclidean distance, the more similar they are), and the second similarity is ranked as 2 in the second similarity ranking result (the larger the Pearson correlation coefficient, the more similar they are). The historical date is calculated according to formula (2). With the date to be predicted The overall similarity between them.

[0058] (2); In equation (2), Indicates historical dates With the date to be predicted Overall similarity between them; Indicates historical dates With the date to be predicted The first similarity ranking results between them; Indicates historical dates With the date to be predicted The second similarity ranking results between them; , These represent the first weighting coefficient and the second weighting coefficient, respectively. Exemplary , The values ​​are 0.75 and 0.25 respectively.

[0059] Following the method described above, calculate other historical dates and the date to be predicted. The overall similarity between the two pairs is calculated, and then they are sorted from smallest to largest to obtain the overall similarity ranking result.

[0060] In some embodiments, a target training dataset is selected from the historical multi-dimensional feature dataset based on the comprehensive similarity corresponding to each historical date, including: The comprehensive similarity corresponding to each historical date is used to select the first candidate training dataset from the historical multi-dimensional feature dataset; The first candidate training dataset is filtered based on the preset first filtering rule to obtain the second candidate training dataset; The second candidate training dataset is selected based on the preset second screening rule to obtain the target training dataset.

[0061] The first screening rule can be based on the 3σ principle (Laida criterion) to remove outliers (outliers) from the first candidate training dataset. The 3σ principle defines the range of outliers and removes them to ensure the reliability of subsequent calculations and judgments.

[0062] The second screening rule is based on cosine similarity. It calculates the mean cosine similarity between each sample in the second candidate training dataset and other samples, and filters out samples with a mean lower than the threshold to achieve intra-sample similarity verification, further purifying the samples and improving sample quality.

[0063] To facilitate understanding, continuing with the previous example, we select the historical multi-dimensional feature data corresponding to the top 25 historical dates ranked in the comprehensive similarity ranking results to form the first candidate training dataset. Next, outliers (outliers) in the first candidate training dataset are filtered out based on the 3σ principle to obtain the second candidate training dataset. The pre-cosine similarity between each pair of second candidate training data points within the second candidate training dataset is calculated. If the mean cosine similarity of a second candidate training data point with other second candidate training data points is lower than a preset threshold (which can be flexibly set according to actual conditions), then that second candidate training data point is discarded. Through the above filtering process, the target training dataset is obtained.

[0064] By adopting the above methods, the overall quality of training samples can be improved, while the robustness of the model training process can be enhanced, ensuring that the prediction model still has stable performance and generalization ability in complex scenarios such as power market supply and demand fluctuations and sudden weather changes.

[0065] In some embodiments, training an initial medium- to long-term price prediction model using a target training dataset includes: The core parameters to be optimized in the initial medium- and long-term price forecasting model are determined, and a standardized parameter optimization network is set up based on the core parameters to be optimized. The target training dataset is divided into a training set and a validation set; With minimizing the root mean square error of the validation set as the optimization objective and the training and validation sets as inputs, the optimal parameter combination is searched within the scope of the standardized parameter optimization network. The model parameters of the initial medium- and long-term price forecasting model are updated using the optimal parameter combination. The updated initial medium- and long-term price forecasting model is trained using the training set as input until the preset convergence condition is met, resulting in a well-trained medium- and long-term price forecasting model.

[0066] The core parameters to be optimized mentioned above include learning rate, maximum tree depth, number of decision trees, regularization coefficient, feature sampling rate, and sample sampling rate.

[0067] As an example, the initial medium- to long-term price prediction model can be a machine learning model such as the XGBoost algorithm. First, the core optimization parameters of the initial medium- to long-term price prediction model (such as the XGBoost algorithm model) are set as follows: learning rate 0.05~0.3, maximum tree depth 3~8, number of decision trees 100~600, regularization coefficient L1=0.1~0.5, feature sampling rate 0.5~1, and sample sampling rate 0.5~1. A standardized parameter optimization network is then set based on these core optimization parameter ranges. Next, the target training dataset is divided into a training set and a validation set according to a preset ratio (e.g., 8:2). The RMSE (root mean square error) of the validation set prediction results is used as the loss function. The core optimization parameters are iteratively trained using a Bayesian optimization algorithm (the number of training iterations can be set to 50, 100, etc.) to search for the optimal parameter combination within the standardized parameter optimization network range. Finally, the optimal parameter combination is used to update the model parameters of the initial medium- and long-term price forecast model. The updated initial medium- and long-term price forecast model is trained using the training set as input until the preset convergence condition is met, resulting in a trained medium- and long-term price forecast model.

[0068] By employing Bayesian optimization algorithms, automated hyperparameter optimization is performed within a parameter space specifically designed to enhance model generalization ability (such as controlling decision tree depth and introducing L1 regularization coefficients). This entire model training process ensures that the final model not only fits well to common historical operating patterns but also possesses excellent inference (generalization) capabilities for complex operating scenarios that are not previously encountered but are reasonable, guaranteeing stable and reliable prediction results.

[0069] In some embodiments, after inputting the multi-dimensional feature data corresponding to the date to be predicted into the trained medium- to long-term price prediction model and outputting the medium- to long-term price prediction result, the method further includes: Input the multi-dimensional feature data corresponding to the date to be predicted into the trained error correction model, and output the prediction residual value corresponding to the date to be predicted; The medium- and long-term price forecast results are corrected based on the predicted residual values ​​to obtain the final medium- and long-term price forecast results.

[0070] As an example, after the basic prediction model has been fully trained and rolling prediction samples have been accumulated, the following error correction prediction process is executed: First, based on the accumulated training sample set, the true mid-to-long-term price value corresponding to each sample is extracted, and the mid-to-long-term price prediction value output by the basic prediction model is calculated. The difference between the two is used as the prediction residual. Then, using the full feature set after feature engineering (i.e., the target training dataset) as input features, and the prediction residual calculated above as the model output target, an error correction model based on gradient boosting trees is constructed. The model training parameters are fixed as follows: tree depth max_depth=3, learning rate learning_rate=0.05, number of decision trees n_estimators=400, and other parameters use default configurations. The error correction model is trained based on the training sample set. Finally, the feature engineering results of the multi-dimensional feature data corresponding to the date to be predicted are input into the trained error correction model to obtain the residual prediction value. The mid-to-long-term price prediction result output by the basic prediction model is added to the residual prediction value to obtain the final mid-to-long-term price prediction result.

[0071] In addition to the core prediction model, this application innovatively introduces an independent error correction model. This model uses the historical prediction residuals of the core prediction model as the learning target, and by mining the temporal distribution patterns and feature correlation patterns of the residuals, it automatically identifies and corrects the systematic biases or inherent errors that may exist in the core model under different operating scenarios.

[0072] The core prediction model and the error correction model work together to form a closed-loop prediction system with self-improvement capabilities. This system not only provides high-precision initial prediction results, but also continuously iterates and optimizes model parameters and correction strategies by accumulating historical operating data, achieving continuous evolution of prediction performance, maintaining long-term technological leadership, and providing stable and reliable technical support for power market participants.

[0073] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0074] In summary, the beneficial effects of the technical solutions provided in this application include at least the following: (1) This application innovatively proposes a cascaded prediction framework of "predicting the boundary first, then predicting the price," which decouples the task of predicting medium- and long-term electricity prices into two relatively independent and easy-to-model sub-tasks. Specifically, it is executed in two stages: In the first stage, using meteorological and environmental characteristic data and time series characteristic data that can be obtained stably over a long period of time, the market supply and demand boundary conditions for the predicted date are accurately calculated through feature selection and a dedicated prediction model. This provides complete and high-quality core input features for the subsequent price prediction model, effectively solving the technical pain point of existing technologies in the process of predicting medium- and long-term electricity prices, which is due to the inability to obtain key power boundary information such as grid load and new energy output for the predicted date in advance, resulting in a lack of core input for the prediction model and difficulty in guaranteeing prediction accuracy. In the second stage, based on the power boundary feature prediction data output from the first stage, combined with the aforementioned historical multi-dimensional feature dataset and the target training dataset after similar day selection, the medium- and long-term electricity price prediction for the predicted date is completed. This two-level prediction task decoupling strategy effectively reduces the modeling complexity and overfitting risk of the overall prediction model, and significantly enhances the interpretability and operational stability of the prediction system.

[0075] By employing the aforementioned two-stage cascaded design, the shortcomings of existing technologies that construct a single "black box" model (i.e., directly mapping long-term meteorological characteristics to electricity prices) are effectively avoided, such as complex model structure, susceptibility to overfitting, and weak generalization ability. At the same time, the cascaded framework has a clear structure and coherent logic, which significantly improves the overall generalization ability, operational stability, and interpretability of the prediction model, effectively overcoming the limitations of existing medium- and long-term electricity price prediction methods.

[0076] (2) This application proposes a dynamic matching technique for similar days based on simulated future scenarios. Using the "bidding space" feature vector output from the first stage, which reflects the expected supply and demand situation on the date to be predicted, as a benchmark, the technique dynamically selects historical multi-dimensional feature data corresponding to the historical date most similar to the supply and demand scenario of the date to be predicted from the historical multi-dimensional feature dataset after outlier processing according to the 3σ principle, and constructs the target training dataset. This technique achieves accurate alignment between the training data and the prediction target at the level of future supply and demand scenarios, providing high-quality, highly relevant sample support for subsequent prediction model training, and further ensuring prediction accuracy.

[0077] The "bidding space" feature vector used for similar day selection is not generated from historical real power boundary data, but rather based on the power boundary feature information of the predicted date (future day) obtained from the first stage of prediction, calculated through linear supply and demand. This design ensures that the selected historical similar days are highly consistent with the market supply and demand scenario "expected" for the predicted date. This similar day matching method based on "simulating future scenarios" ensures that the target training dataset and the prediction target resonate with each other in terms of fundamental market supply and demand drivers. This allows the historical price change patterns learned by the prediction model to be directly adapted and applied to the specific supply and demand scenario of the predicted date, thereby significantly improving the targeting, accuracy, and reliability of medium- and long-term power price prediction.

[0078] (3) This application proposes a meteorological-electricity correlation feature construction and screening technology for boundary information prediction. In order to achieve accurate prediction of various types of electricity boundary information, a standardized feature screening method is designed: from multi-dimensional meteorological features and time series features, by calculating the Pearson correlation coefficient between each feature and specific electricity boundary information, a subset of strongly correlated features with an absolute value of Pearson correlation coefficient greater than a preset threshold is automatically identified and screened. Based on this feature subset, a dedicated simple XGBoost prediction model is constructed to ensure the reliability and accuracy of electricity boundary information prediction results and provide a solid data foundation for subsequent price prediction.

[0079] (4) In view of the characteristics of long-term electricity price forecasting, complex price fluctuation patterns, and potentially limited but highly correlated training samples, this application designs a comprehensive adaptation and optimization scheme, including a meteorological utility reconstruction method based on new energy installed capacity and a time feature encoding method oriented towards long-term trends. Simultaneously, for the limited and highly correlated target training samples obtained after dynamic matching and screening of similar days (such as historical samples ranked in the top 25 in comprehensive similarity), a lightweight XGBoost model parameter space and Bayesian optimization strategy are specifically designed to ensure that the prediction model can converge quickly under limited sample conditions and maintain strong generalization ability. In summary, through the organic combination of the above technologies, this application ultimately forms a complete, efficient, and robust industrial-grade solution for medium- and long-term electricity price forecasting, which can be widely applied to medium- and long-term electricity market price forecasting scenarios.

[0080] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0081] Figure 2 This is a schematic diagram of a medium- to long-term price forecasting device provided in an embodiment of this application. Figure 2 As shown, the medium- to long-term price forecasting device 200 includes: The data acquisition module 201 is configured to collect historical multi-dimensional feature datasets by tracing back from the date to be predicted. The historical multi-dimensional feature datasets include multiple historical multi-dimensional feature data, and each historical multi-dimensional feature data includes power system operation feature data, meteorological environment feature data, and power market transaction feature data for its corresponding historical date. The determination module 202 is configured to determine the power boundary feature data corresponding to the date to be predicted based on a historical multi-dimensional feature dataset. The filtering module 203 is configured to filter out the target training dataset from the historical multi-dimensional feature dataset based on the power boundary feature data and the historical multi-dimensional feature dataset. The first training module 204 is configured to train an initial medium- and long-term price prediction model using the target training dataset until a preset convergence condition is met, thereby obtaining a trained medium- and long-term price prediction model. The prediction module 205 is configured to input the multi-dimensional feature data corresponding to the date to be predicted into the trained medium- and long-term price prediction model and output the medium- and long-term price prediction results.

[0082] In some embodiments, the determining module 202 includes: The first determining unit is configured to determine the historical power boundary feature data corresponding to the historical date of each historical multi-dimensional feature data; The first filtering unit is configured to filter the corresponding historical multi-dimensional feature data based on historical power boundary feature data to obtain the target multi-dimensional feature data. The second determining unit is configured to determine the power boundary feature data corresponding to the date to be predicted based on the target multi-dimensional feature data.

[0083] In some embodiments, the filtering module 203 described above includes: The third determining unit is configured to determine the target bidding space vector based on power boundary feature data and to determine the historical bidding space vector set based on historical multi-dimensional feature dataset. The second filtering unit is configured to filter out the target training dataset from the historical multi-dimensional feature dataset based on the target bidding space vector and the historical bidding space vector set.

[0084] In some embodiments, the historical bidding space vector set includes multiple historical bidding space vectors, and one historical bidding space vector corresponds to one historical date. The second filtering unit described above includes: The first calculation component is configured to calculate the first similarity and the second similarity between the target bidding space vector and each historical bidding space vector; The second calculation component is configured to calculate the comprehensive similarity between the historical date and the date to be predicted for each historical date, based on the first similarity and the second similarity corresponding to the historical date. The filtering component is configured to select the target training dataset from the historical multi-dimensional feature dataset based on the comprehensive similarity corresponding to each historical date.

[0085] In some embodiments, the filtering component described above may be specifically configured as follows: The comprehensive similarity corresponding to each historical date is used to select the first candidate training dataset from the historical multi-dimensional feature dataset; The first candidate training dataset is filtered based on the preset first filtering rule to obtain the second candidate training dataset; The second candidate training dataset is selected based on the preset second screening rule to obtain the target training dataset.

[0086] In some embodiments, the training module 204 described above may be specifically configured as follows: The core parameters to be optimized in the initial medium- and long-term price forecasting model are determined, and a standardized parameter optimization network is set based on the core parameters to be optimized. The core parameters to be optimized include learning rate, maximum tree depth, number of decision trees, regularization coefficient, feature sampling rate, and sample sampling rate. The target training dataset is divided into a training set and a validation set; With minimizing the root mean square error of the validation set as the optimization objective and the training and validation sets as inputs, the optimal parameter combination is searched within the scope of the standardized parameter optimization network. The model parameters of the initial medium- and long-term price forecasting model are updated using the optimal parameter combination. The updated initial medium- and long-term price forecasting model is trained using the training set as input until the preset convergence condition is met, resulting in a well-trained medium- and long-term price forecasting model.

[0087] In some embodiments, the above-described apparatus further includes: The second training module is configured to input the multi-dimensional feature data corresponding to the date to be predicted into the trained error correction model and output the prediction residual value corresponding to the date to be predicted. The correction module is configured to correct the medium- and long-term price forecast results based on the forecast residual values ​​to obtain the final medium- and long-term price forecast results.

[0088] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0089] Figure 3 This is a schematic diagram of the electronic device 300 provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device 300 of this embodiment includes a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various device embodiments described above.

[0090] Electronic device 300 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 300 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 300 and does not constitute a limitation on electronic device 300. It may include more or fewer parts than shown, or different parts.

[0091] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0092] The memory 302 can be an internal storage unit of the electronic device 300, such as a hard disk or RAM of the electronic device 300. The memory 302 can also be an external storage device of the electronic device 300, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 300. The memory 302 can also include both internal and external storage units of the electronic device 300. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0093] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0094] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium can be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, a computer-readable medium does not include electrical carrier signals and electrical signals.

[0095] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A medium- to long-term price forecasting method, characterized in that, include: Starting from the date to be predicted, historical multi-dimensional feature datasets are collected backward. The historical multi-dimensional feature datasets include multiple historical multi-dimensional feature data, and each historical multi-dimensional feature data includes power system operation feature data, meteorological environment feature data, and power market transaction feature data for its corresponding historical date. Based on the historical multi-dimensional feature dataset, determine the power boundary feature data corresponding to the date to be predicted; Based on the power boundary feature data and the historical multi-dimensional feature dataset, a target training dataset is selected from the historical multi-dimensional feature dataset. The initial medium- and long-term price prediction model is trained using the target training dataset until the preset convergence condition is met, thus obtaining the trained medium- and long-term price prediction model. The multi-dimensional feature data corresponding to the date to be predicted is input into the trained medium- and long-term price prediction model, and the medium- and long-term price prediction results are output.

2. The method according to claim 1, characterized in that, Based on the historical multi-dimensional feature dataset, the power boundary feature data corresponding to the date to be predicted is determined, including: Determine the historical power boundary feature data for each historical date corresponding to the aforementioned historical multi-dimensional feature data; Based on the historical power boundary feature data, the corresponding historical multi-dimensional feature data is filtered to obtain the target multi-dimensional feature data; Based on the target multi-dimensional feature data, the power boundary feature data corresponding to the date to be predicted is determined.

3. The method according to claim 1, characterized in that, Based on the aforementioned power boundary feature data and historical multi-dimensional feature dataset, a target training dataset is selected from the historical multi-dimensional feature dataset, including: The target bidding space vector is determined based on the power boundary feature data, and the historical bidding space vector set is determined based on the historical multi-dimensional feature dataset. Based on the target bidding space vector and the historical bidding space vector set, the target training dataset is selected from the historical multi-dimensional feature dataset.

4. The method according to claim 3, characterized in that, The historical bidding space vector set includes multiple historical bidding space vectors, and each historical bidding space vector corresponds to a historical date; Based on the target bidding space vector and the historical bidding space vector set, a target training dataset is selected from the historical multi-dimensional feature dataset, including: Calculate the first similarity and the second similarity between the target bidding space vector and each of the historical bidding space vectors; For each historical date, a comprehensive similarity between the historical date and the date to be predicted is calculated based on the first similarity and the second similarity corresponding to the historical date. Based on the comprehensive similarity corresponding to each of the historical dates, the target training dataset is selected from the historical multi-dimensional feature dataset.

5. The method according to claim 4, characterized in that, Based on the comprehensive similarity corresponding to each of the aforementioned historical dates, a target training dataset is selected from the historical multi-dimensional feature dataset, including: The comprehensive similarity corresponding to each of the historical dates is used to select the first candidate training dataset from the historical multi-dimensional feature dataset; The first candidate training dataset is filtered based on a preset first filtering rule to obtain a second candidate training dataset; The second candidate training dataset is filtered based on the preset second filtering rule to obtain the target training dataset.

6. The method according to claim 1, characterized in that, Training an initial medium- to long-term price prediction model using the target training dataset includes: The core parameters to be optimized in the initial medium- and long-term price forecasting model are determined, and a standardized parameter optimization network is set based on the core parameters to be optimized. The core parameters to be optimized include learning rate, maximum tree depth, number of decision trees, regularization coefficient, feature sampling rate, and sample sampling rate. The target training dataset is divided into a training set and a validation set; With minimizing the root mean square error of the validation set as the optimization objective and the training set and validation set as input, the optimal parameter combination is searched within the scope of the standardized parameter optimization network. The model parameters of the initial medium- and long-term price prediction model are updated using the optimal parameter combination. The updated initial medium- and long-term price prediction model is trained using the training set as input until the preset convergence condition is met, and the trained medium- and long-term price prediction model is obtained.

7. The method according to claim 1, characterized in that, After inputting the multi-dimensional feature data corresponding to the date to be predicted into the trained medium- and long-term price prediction model and outputting the medium- and long-term price prediction results, the model further includes: Input the multi-dimensional feature data corresponding to the date to be predicted into the trained error correction model, and output the prediction residual value corresponding to the date to be predicted; The medium- and long-term price forecast results are corrected based on the predicted residual values ​​to obtain the final medium- and long-term price forecast results.

8. A medium- to long-term price forecasting device, characterized in that, include: The data acquisition module is configured to collect historical multi-dimensional feature datasets by tracing back from the date to be predicted. The historical multi-dimensional feature datasets include multiple historical multi-dimensional feature data, and each historical multi-dimensional feature data includes power system operation feature data, meteorological environment feature data, and power market transaction feature data for its corresponding historical date. The determination module is configured to determine the power boundary feature data corresponding to the date to be predicted based on the historical multi-dimensional feature dataset; The filtering module is configured to filter out the target training dataset from the historical multi-dimensional feature dataset based on the power boundary feature data and the historical multi-dimensional feature dataset. The training module is configured to train an initial medium- to long-term price prediction model using the target training dataset until a preset convergence condition is met, thereby obtaining a trained medium- to long-term price prediction model. The prediction module is configured to input the multi-dimensional feature data corresponding to the date to be predicted into the trained medium- and long-term price prediction model and output the medium- and long-term price prediction results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images