A method and system for real-time nodal electricity price forecasting based on multi-source time series fusion and meteorological adaptive correction
By employing a multi-source time-series fusion and meteorological adaptive correction method, and combining a residual LSTM model with a long-term macroeconomic base, the electricity price prediction method solves the problems of time misalignment and scenario adaptation in multi-source power data processing, and achieves high-precision real-time electricity price prediction.
Patent Information
- Application Number
- CN202611107515.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing real-time electricity price forecasting methods suffer from problems such as time misalignment, feature leakage, inability of models to balance near-term forecast accuracy and long-term stability, complex influence of meteorological factors, and lack of scenario adaptation mechanisms when processing multi-source electricity data, leading to distorted forecast results.
By employing a multi-source time-series fusion and meteorological adaptive correction method, and through dual-track parallel prediction of residual LSTM model and long-term macroeconomic base, combined with dynamic gating and post-form correction, we can achieve leakage-free, near-end sensitive and long-term stable electricity price prediction.
It achieves leak-free, near-end sensitive, and long-term stable real-time electricity price forecasting for nodes, improving the production reliability and historical backtesting accuracy of the forecasting system, and demonstrating excellent performance adaptable to different forecast periods and operating scenarios.
Smart Images

Figure CN123121604A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of real-time electricity price forecasting technology in the electricity market, and more specifically, to a method and system for real-time nodal electricity price forecasting based on multi-source time-series fusion and meteorological adaptive correction. Background Technology
[0002] In a market-based electricity trading environment, real-time electricity prices at nodes are influenced by multiple factors, including load levels, generating unit output, renewable energy penetration rates, non-market generating unit plans, hydropower and pumped storage, weather conditions, holiday effects, and market boundary conditions. These factors result in high volatility, strong cyclicality, periods of negative prices, and sudden price changes. Real-time electricity price forecasting is crucial for electricity trading decisions, load management, energy storage dispatch, virtual power plant operation, and risk monitoring. Existing forecasting methods typically employ single time-series models such as ARIMA or single machine learning models such as support vector machines or random forests. Some solutions combine similar day retrieval with LSTM neural networks or use multiple models for simple weighted averaging. However, these methods generally suffer from the following shortcomings.
[0003] First, multi-source power operation data have different time granularities, different business definitions, and different release time rules. Real-time electricity prices are usually released at 15-minute or hourly granularities, day-ahead prices are released at fixed times every day, and generation forecasts, renewable energy forecasts, and load forecasts each have different forecast durations and update times, while meteorological data comes from independent forecasting systems. Existing technologies often directly splice or simply align these data, which can easily lead to time misalignment, feature leakage, or inconsistent definitions. Especially during backtesting, if the visibility boundaries of the data are not strictly controlled, future data may be unintentionally used, resulting in distorted backtesting results that cannot truly reflect online forecasting performance. In addition, a single forecasting model cannot simultaneously achieve both near-term forecast accuracy and long-term forecast stability. Neural network models such as LSTM can obtain high sensitivity in the near term by utilizing recent prices and operational characteristics, but when extrapolating to the future over multiple days, they are prone to oscillations due to baseline missing, forecast feature errors, and distribution drift; similar day and template methods have better long-term stability, but they are insufficient in responding to short-term boundary changes, high renewable energy penetration, and extreme weather. While existing technologies offer solutions for fusing multiple models, they typically employ static weighting or black-box models, such as GBDT (Guided Graph Theory) which learns fusion weights. These methods cannot adaptively adjust the fusion strategy based on scenario factors like the number of forecast days, the proportion of renewable energy, and operational fluctuations, and they also struggle to explain the physical meaning of weight changes. Furthermore, the impact of meteorological factors on real-time electricity prices is complex and significantly scenario-dependent. Strong sunshine may lead to increased renewable energy output and lower electricity prices at midday; thick clouds or rainfall can weaken this effect. Strong winds at night may affect wind power output and off-peak prices, while high or low temperatures may increase evening peak load and prices. Existing technologies often use meteorological variables as pre-input features in the model, which mixes meteorological signals with strong time-series signals such as load and electricity prices, making it difficult for the model to specifically learn the impact of meteorology on intraday price patterns. Moreover, electricity consumption structure, unit combinations, and market trading behavior change during holidays and weekends, significantly altering the transmission relationship from meteorology to electricity prices compared to ordinary workdays. Existing technologies lack targeted scenario adaptation mechanisms, making them prone to prediction bias during long holidays due to applying ordinary daily patterns.
[0004] Therefore, researching and designing a real-time electricity price forecasting method and system based on multi-source time-series fusion and meteorological adaptive correction that can overcome the above-mentioned defects is an urgent problem to be solved. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for real-time nodal electricity price forecasting based on multi-source time-series fusion and meteorological adaptive correction. First, strict visibility isolation eliminates future data leakage, ensuring that backtesting and online forecasting are in the same information state. Then, a dual-track parallel approach using residual LSTM and a long-term baseline is employed, respectively responsible for near-end sensitive response and long-term stable pattern. Next, dynamic gating is used to adaptively switch the dominant force based on the number of forecast days, the proportion of renewable energy, and net load fluctuations. Finally, a scene-adaptive pattern-level correction is applied to the fusion results using a meteorological pattern library, holiday weakening factors, and negative price release rules. This achieves leak-free, near-end sensitive, long-term stable, meteorological adaptive, and backtestable real-time nodal electricity price forecasting.
[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: Firstly, a nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction is provided, including the following steps: Acquire multi-source power data, which includes at least historical real-time electricity prices, day-ahead prices, total power generation output forecasts, new energy power generation output forecasts, load forecasts, historical meteorological data, and future meteorological data; The multi-source power data is standardized into a unified time grid, and the real-time electricity price target column and the new energy characteristic caliber are unified. Using the decision-making time point as the boundary, a leak-free feature set is constructed, wherein the real-time electricity price after the decision-making time point and the future invisible features are masked, and only the feature data published before the decision-making time point is retained; Based on the leak-free feature set, a residual LSTM model is trained, with the residual of the real-time electricity price relative to the baseline price as the learning target. Construct a forward macroeconomic foundation, which includes a calendar fractal template based on historical prices, an exponential smoothing extreme value extrapolation curve, and a similar day price curve; The prediction results of the residual LSTM model are fused with the long-term macroeconomic base through dynamic gating, wherein the gating weights are adaptively calculated based on the number of prediction days, the proportion of new energy sources, and net load fluctuations. Based on historical meteorological price pattern database and future meteorological data, a post-form-level correction is performed on the fused forecast results, and combined with holiday weakening factors and negative price release rules, the final forecast results are obtained.
[0007] Furthermore, the construction of the leak-free feature set includes: The decision-making time point simultaneously constrains the available data boundaries of the training set, calendar classification template library, similar day historical sample library, and historical meteorological price pattern library; The actual real-time electricity price after the decision point is only used as a backtesting label or post-evaluation indicator and does not enter any training, template construction, similar day retrieval or morphology library calibration process. The future invisible features include day-ahead prices and daily forecasts that have not yet been published after the decision-making time.
[0008] Furthermore, the definition of the new energy characteristics is standardized, including: The effective total new energy volume is selected in descending order of business priority, namely, the sum of the new energy sub-type columns, the total new energy volume column, the daily new energy forecast column, and the weekly new energy forecast column. Furthermore, low-priority data will not be accumulated when high-priority data is available.
[0009] Furthermore, the baseline price adopts the historical date-ahead price or business benchmark price during the training phase, and is dynamically updated according to the latest visible real-time electricity price during the rolling inference phase. The feature sequence input to the residual LSTM model includes at least the hourly index, weekday index, net load, renewable energy percentage, currently used baseline price, lagged price, and rolling statistics. After training is completed, rolling inference is performed on the prediction interval. During each inference, the baseline price is updated based on the latest real-time electricity price published before the current decision point. The updated baseline price is added to the residual predicted by the model to obtain the predicted price at the current time point. This predicted price is then used as the lagged price input in the feature sequence at the next time point, and the process is repeated point by point to form a continuous 15-minute near-end prediction sequence.
[0010] Furthermore, the construction of the long-term macroeconomic foundation includes: For each calendar type and weekday combination, extract the historical intraday price curve and normalize it to obtain a standard pattern template. An exponential smoothing model is used to predict the price extremes of each intraday segment, and the standard pattern template is stretched to the predicted extreme value range to obtain the template reconstruction curve; Based on the net load or weather signature of the future forecast period, retrieve several similar days from the historical sample, and calculate the average or weighted average of the historical price curves of the similar days to obtain the similar day curves; The template reconstruction curve and the similar day curve are fused together according to a preset weight to obtain the long-term macroeconomic base.
[0011] Furthermore, the weights of the dynamic gating are calculated as follows: The gating weight is jointly determined by the exponential decay term corresponding to the number of forecast days, the positive gain term corresponding to the proportion of new energy, and the positive gain term corresponding to the intraday range of net load, and the gating weight is restricted to the range [0,1]. The larger the number of days to forecast, the smaller the exponential decay term, which increases the weight of the long-term macroeconomic base in the long-term forecast. The higher the proportion of new energy sources or the larger the intraday range of net load, the larger the positive gain term, which increases the weight of the residual LSTM model in near-end prediction. The fusion prediction result is a weighted sum of the prediction results of the residual LSTM model and the long-term macroeconomic base.
[0012] Furthermore, the post-morphological level correction includes: A historical meteorological price pattern database is constructed based on historical meteorological data and historical real-time prices. Each historical natural day in the database contains a weather signature and an intraday price pattern curve. The weather signature includes at least the average noon solar irradiance, average cloud cover, average nighttime wind speed, total precipitation, and average temperature. For a predicted natural day, its weather signature is calculated, similar weather days are retrieved from the morphology library, and a weather prototype curve is constructed. Based on one or more of the following factors: irradiance, cloud cover, rainfall, nighttime wind intensity, and heat or cold load intensity, directional corrections are applied to the intraday price pattern of the fused forecast results. These directional corrections include lowering midday prices during periods of strong sunshine, reducing the midday price drop during periods of thick clouds or rainfall, adjusting trough prices during periods of strong nighttime winds, and raising evening peak prices during periods of high or low temperatures.
[0013] Furthermore, the holiday weakening factor takes a first weakening coefficient during the Spring Festival and National Day holidays, a second weakening coefficient during the pre-holiday and post-holiday windows, a third weakening coefficient during ordinary weekends, and a fourth weakening coefficient during adjusted workdays. The first weakening coefficient is less than the second weakening coefficient, the second weakening coefficient is less than the third weakening coefficient, and the third weakening coefficient is less than the fourth weakening coefficient. The negative price release rule is triggered when strong sunlight, a high proportion of new energy sources, and low net load all meet preset conditions simultaneously, causing the price during the midday period to be released towards the historical negative price percentile, and is constrained by the lower limit of the market price.
[0014] Furthermore, the standardization of the multi-source power data into a unified time grid includes: The date field and the time field are combined into a timestamp index. When the value of the time field is 24:00, it is converted to the next day's midnight as the modeling index, and in the final output, it is restored to the 24:00 record of the previous trading day. The final prediction results are aggregated by hourly right-closed intervals and written to the database. Before writing, old prediction records within the target date range are deleted.
[0015] Secondly, a real-time nodal electricity price forecasting system based on multi-source time-series fusion and meteorological adaptive correction is provided, including: The data acquisition module is used to acquire multi-source power data, which includes at least historical real-time electricity prices, day-ahead prices, total power generation output forecasts, new energy output forecasts, load forecasts, historical meteorological data, and future meteorological data. The unification module is used to standardize the multi-source power data into a unified time grid and unify the real-time electricity price target column and the new energy characteristic caliber. The feature construction module is used to construct a leak-free feature set with the decision time point as the boundary. The real-time electricity price after the decision time point and the future invisible features are masked, and only the feature data published before the decision time point is retained. The model training module is used to train a residual LSTM model based on the leak-free feature set, wherein the residual LSTM model uses the residual of the real-time electricity price relative to the baseline price as the learning target. The base construction module is used to construct a forward macroeconomic base, which includes a calendar fractal template based on historical prices, an exponential smoothing extreme value extrapolation curve, and a similar day price curve. The gating fusion module is used to dynamically gating and fusing the prediction results of the residual LSTM model with the long-term macroeconomic base, wherein the gating weights are adaptively calculated based on the number of prediction days, the proportion of new energy sources, and net load fluctuations. The final prediction module is used to perform post-form-level corrections on the fused prediction results based on the historical meteorological price pattern database and future meteorological data, and to obtain the final prediction results by combining the holiday weakening factor and the negative price release rule.
[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction provided by this invention constructs a complete closed-loop link from data access to final prediction by acquiring multi-source power data and sequentially performing steps such as time standardization and caliber unification, visibility isolation of decision-making time points, residual LSTM near-end prediction, construction of a long-term macroeconomic base, dynamic gating fusion, and post-meteorological morphology correction. First, strict visibility isolation eliminates future data leakage, ensuring that backtesting and online prediction are in the same information state. Then, the residual LSTM and the long-term base are used in parallel, respectively responsible for near-end sensitive response and long-term stable morphology. Next, dynamic gating is used to adaptively switch the dominant position based on the number of prediction days, the proportion of renewable energy, and net load fluctuations. Finally, the fusion results are modified at the morphological level to adapt to the scenario using a meteorological morphology library, holiday weakening factors, and negative price release rules. This invention achieves no leakage, near-end sensitivity, long-term stability, meteorological adaptation, and backtesting verifiable nodal real-time electricity price forecasting.
[0017] 2. When constructing a leak-free feature set, this invention simultaneously constrains the available data boundaries of the training set, calendar classification template library, similar day historical sample library, and historical meteorological price pattern library at the decision-making time point, ensuring that all sub-modules operate under the same information state. Simultaneously, by selecting the sum of the new energy sub-type columns, the total amount column, the daily forecast column, and the weekly forecast column as the effective new energy total amount according to business priority, it avoids net load jumps caused by field caliber switching. This invention eliminates information asymmetry between backtesting and online forecasting from the data source and ensures the long-term stability of feature caliber, thereby improving the authenticity of historical backtesting evaluations and the production reliability of the forecasting system.
[0018] 3. This invention uses a residual LSTM model with the residual between the real-time electricity price and the baseline price as the learning target. During the rolling inference phase, the baseline price is dynamically updated based on the latest available real-time electricity price, ensuring the model consistently focuses on learning incremental changes in market conditions rather than absolute price levels. Simultaneously, a long-term macroeconomic foundation is constructed through calendar fractal templates, exponential smoothing extreme value extrapolation, and similar-day KNN fusion, providing a stable and physically plausible price pattern for long-term forecasting. This invention solves the oscillation problem caused by baseline loss and distribution drift in long-term extrapolation of single neural network models, achieving an organic unity between sensitive near-term response to operational changes and stable control of price patterns in the long term.
[0019] 4. This invention utilizes dynamic gating weights, jointly determined by the exponential decay term corresponding to the number of forecast days, the positive gain term corresponding to the proportion of renewable energy, and the positive gain term corresponding to the intraday range of net load. This allows the fusion weights to adaptively adjust with the forecast period and operating scenario: the longer the forecast period, the higher the weight of the long-term base model; the higher the proportion of renewable energy or the greater the fluctuation of net load, the higher the weight of the short-term model. This invention overcomes the shortcomings of traditional static weighting, which cannot adapt to different forecast periods and scenarios. This invention maintains excellent forecasting performance under different time scales and market conditions.
[0020] 5. This invention constructs a historical meteorological price pattern database, calculates weather signatures for predicted natural days, and retrieves similar weather days to construct weather prototype curves. It then performs directional pattern correction on the fused prediction results. Simultaneously, it sets different weakening coefficients for long holidays, before and after holidays, weekends, and adjusted workdays through holiday weakening factors. Furthermore, it triggers negative price release under conditions of strong sunshine, high renewable energy, and low net load, subject to a price lower limit constraint. This invention transforms meteorological data from a conventional pre-input feature into a post-processing pattern-level correction, and overlays holiday scenario adaptation and safety boundary constraints. This solves the problems of existing technologies where meteorological data is only used as a static feature, misuse of ordinary daily patterns during holidays, and inability to reasonably release negative price risks. This makes the prediction curve both sensitive and robust under high renewable energy weather scenarios.
[0021] 6. This invention combines date and time fields into a timestamp index and performs modeling transformation and output restoration for 24-hour intervals, ensuring the continuity of the 15-minute unified time grid and the compatibility of business expressions. The final prediction results are aggregated by hourly right-closed intervals and written to the database. Before writing, old prediction records within the target date range are deleted to ensure data consistency and traceability. Simultaneously, through the sequential connection of modules such as data access, time standardization, unified caliber, visibility isolation, residual prediction, foundation construction, dynamic gating, meteorological correction, and diagnostic backtesting, a complete closed loop is formed from data to prediction to production application. Attached Figure Description
[0022] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart from Embodiment 1 of the present invention; Figure 2 This is a prediction effect diagram from Embodiment 1 of the present invention; Figure 3 This is a system block diagram in Embodiment 2 of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.
[0024] Example 1: A nodal real-time electricity price forecasting method based on multi-source time series fusion and meteorological adaptive correction, such as Figure 1 As shown, it includes the following steps: S1: Obtain multi-source power data, which includes at least historical real-time electricity prices, day-ahead prices, total power generation output forecasts, new energy power generation output forecasts, load forecasts, historical meteorological data, and future meteorological data; S2: Standardize multi-source power data into a unified time grid, and unify the real-time electricity price target column and the new energy characteristic caliber; S3: Construct a leak-free feature set with the decision time point as the boundary, in which the real-time electricity price after the decision time point and future invisible features are masked, and only the feature data published before the decision time point is retained; S4: Based on the leak-free feature set, train a residual LSTM model. The residual LSTM model uses the residual of the real-time electricity price relative to the baseline price as the learning target. S5: Construct a long-term macroeconomic foundation, which includes a calendar fractal template based on historical prices, an exponential smoothing extreme value extrapolation curve, and a similar day price curve; S6: The prediction results of the residual LSTM model are fused with the long-term macroeconomic foundation through dynamic gating, where the gating weights are adaptively calculated based on the number of prediction days, the proportion of new energy sources, and net load fluctuations. S7: Based on the historical meteorological price pattern database and future meteorological data, a post-form-level correction is performed on the fused prediction results, and the final prediction results are obtained by combining the holiday weakening factor and the negative price release rule.
[0025] In step S1, multi-source power data is first acquired. This data includes at least historical real-time electricity prices, day-ahead prices, total power generation output forecasts, renewable energy output forecasts, load forecasts, historical meteorological data, and future meteorological data. Historical real-time electricity prices refer to the actual settlement price at the node or bus in the electricity market, typically published in 15-minute or hourly increments. Day-ahead prices refer to the clearing prices for each time period of the following day published daily by the trading center. Total power generation output forecasts refer to the expected total output of all generating units in the entire network within a future time period, published by the grid dispatching department. Renewable energy output forecasts include short-term and ultra-short-term power forecasts for wind and solar power. Load forecasts refer to the short-term forecast values of the system's total load. Historical meteorological data includes observed values of temperature, wind speed, cloud cover, irradiance, and precipitation over a past period. Future meteorological data consists of hourly forecast values of meteorological elements provided by meteorological forecasting departments.
[0026] The aforementioned multi-source data can originate from the same database or different databases, and can also access remote databases via secure tunnels. During data access, the availability of each database connection is first verified to ensure a normal connection and accessible data tables. Simultaneously, the business filtering conditions, target region, target node, line type, time range, and predicted start and end dates for each data table are recorded. In actual deployment, sensitive connection parameters such as database usernames, passwords, hosts, and ports are managed by a security configuration module or environment variables and are not considered essential limitations of this invention.
[0027] All the acquired raw data will be organized into a raw dataset. This dataset contains multiple business fields, including at least the date field (ts_day), time point field (time_point), quantity field, data type field (type), release time field (forecast_time), province field (province), node field (node), and line type field (line_type). The date field indicates the calendar day to which the data belongs, the time point field indicates the specific time within that calendar day, and the combination of the two uniquely identifies a 15-minute or hourly time point; the value field is the corresponding numerical value at that time point; the data type field is used to distinguish different data categories, such as real-time electricity prices, day-ahead prices, wind power forecasts, and photovoltaic forecasts; the release time field indicates the release time of the data, used for subsequent data visibility determination. This raw dataset will be input into the next step for time standardization and feature concatenation.
[0028] In step S2, the date and time fields from the obtained raw dataset are first combined into a timestamp index. Specifically, the date field `ts_day` and the time field `time_point` are concatenated into a continuous time index of type `datetime`. When the value of the time field is exactly 24:00, the time of that record is converted to midnight of the next day, and the date field is shifted forward by one day, thus ensuring that midnight belongs to the previous trading day while the time index remains continuous. In the subsequent final output stage, the midnight record is restored to the 24:00 record of the previous trading day to conform to the business time expression habits of the electricity market.
[0029] For data tables with data type fields, deduplication and pivoting operations are performed using the timestamp index and data type field as keys, converting numeric fields corresponding to different data types into multi-column features. For data tables without data type fields, numeric fields are directly mapped to feature columns corresponding to the business prefix. All features are used to generate a unified skeleton index at a 15-minute frequency and then concatenated into a complete master data table using a left join. For missing features in the master data table, values from the same time the previous day are used for imputation first, followed by values from the same time the previous week, and then forward imputation and conservative zero-value imputation are performed. This imputation strategy utilizes the daily and weekly cycles of electricity load and price while ensuring a complete feature grid for future forecast intervals.
[0030] Next, we unify the real-time electricity price target column. We map the real-time price column of the target node or bus to a unified `rt_price`. For different original column names from the node price table, bus price table, or regional price table, we filter them by target node, line type, region, and service type to form the same target column, ensuring that the same price target is used for training, backtesting, and database entry.
[0031] Next, standardize the definition of renewable energy characteristics. Renewable energy data may exist in various forms in the business system, including sub-categories of wind and solar power, total renewable energy volume, daily renewable energy forecasts, and weekly renewable energy forecasts. To avoid duplicate inclusion of features with different definitions in net load calculations, they are processed in descending order of business priority: the sum of the sub-categories of renewable energy is used as the effective total renewable energy volume first; if the sub-categories are unavailable, the total renewable energy volume column is used; if the total renewable energy volume column is unavailable, the daily renewable energy forecast column is used; if the daily renewable energy forecast column is unavailable, the weekly renewable energy forecast column is used. When high-priority data is available, low-priority data is not accumulated, thus ensuring the stability of net load characteristics.
[0032] In one implementation, the net load characteristic can be obtained by subtracting the total renewable energy output forecast from the total load forecast; in another implementation, when the business uses the total generation forecast as the supply-side benchmark, it can also be obtained by subtracting the total renewable energy output forecast from the total generation output forecast. The formula for calculating the net load characteristic is as follows: ; Or, for: ; in, Indicates a time point of fifteen minutes. This represents the net load characteristics at that point in time. This is the total load forecast value. This is a unified forecast of total new energy power output. This represents the predicted total power generation output. Net load characteristics are used to characterize supply and demand pressure and are important inputs for subsequent similar day retrieval and dynamic gating weight calculation.
[0033] After completing the above processing, the resulting master data table will include at least the real-time electricity price target column `rt_price`, the baseline price column `baseline_price`, the net load characteristic `fcst_net_load`, the weekday index `dayofweek`, the hour index `hour`, the calendar type, the proportion of renewable energy, and optional meteorological characteristics. This master data table will be input into the next step for decision-making visibility isolation.
[0034] In step S3, the decision time point Construct a leak-free feature set for the boundary. The decision point refers to the current moment when prediction is made. For any 15-minute time point... Only data records whose publication time is no later than the current decision-making time are allowed to form feature sets. Publication time refers to the actual time when each data record is available for querying in the business system, denoted as […]. ,in This represents a specific data record. The construction rule for the feature set is expressed by the following formula: ; in, Indicates the point in time of decision-making Next time point Construct a leak-free feature set, This is the main data table output from the previous step. For data recording Release time, Indicates time point Or the set of data records related to its historical window. Only data records whose publication time is no later than the current decision point and which are related to the current prediction point or its historical window are allowed to be included in the feature set.
[0035] Specifically, the timing of the decision Subsequent real-time electricity prices are set to null and are not used as input to any model. Day-ahead prices and daily forecast features after the decision point are only retained within the business visibility range; those outside the visibility range are masked. Weekly forecast features are retained only within the range covered by weekly forecasts. Historical real electricity prices are only used for training before the decision point, calendar pattern template construction, similar day retrieval, and historical weather price pattern database calibration.
[0036] Decision timing Simultaneously, the available data boundaries of the training set, calendar pattern template library, similar day historical sample library, and historical weather price pattern library are constrained. In other words, all historical data used in subsequent steps are based on the decision-making time. Data after the decision-making time is not visible. Actual real-time electricity prices after the decision-making time are only used as backtesting labels or post-event evaluation indicators and are not included in any training, template construction, similar day retrieval, or pattern library calibration processes. Future invisible features include day-ahead prices and daily forecast features that have not yet been published after the decision-making time; these features are not available during the training and inference phases.
[0037] The aforementioned visibility isolation mechanism ensures that the features used in the training set, validation set, and online prediction are in the same information state, avoiding distortion of backtesting results due to future data leakage. After visibility isolation, a leak-free feature set is obtained, with its training interval ending 15 minutes before the decision point and its inference interval covering the prediction start and end points. This leak-free feature set will be input into the training and inference of the residual LSTM model, as well as the construction of the long-term macroscopic foundation.
[0038] In step S4, a residual LSTM model is trained based on the leak-free feature set obtained in the previous step. This model does not directly predict highly volatile real-time electricity prices, but instead constructs a residual relative to the baseline price as the learning objective. The residual objective is defined as the difference between the actual real-time electricity price and the baseline price, calculated as follows: ; in, For time points The residual objective, This represents the actual value of the real-time electricity price. The baseline price is used during the training phase, employing either the historical date-ahead price or a pre-set business benchmark price. During the rolling inference phase, the baseline price is dynamically updated based on the latest real-time electricity price published before the current decision point, ensuring the model continuously learns the deviation relative to the latest market state.
[0039] The feature sequence input to the residual LSTM model includes at least the hourly index, weekday index, net load, renewable energy percentage, the currently used baseline price, and lagged prices and rolling statistics. The historical window length of the feature sequence is L fifteen-minute points, forming the input matrix. The model uses two LSTM layers to temporally encode the feature sequence, extracting the temporal dependencies, and then maps them to single-point residual predictions through fully connected layers and nonlinear activation layers. The model uses the mean absolute error as the loss function and is trained using the Adam optimizer.
[0040] After training, rolling inference is performed on the prediction interval. During each inference, the baseline price is first updated based on the latest real-time electricity price published before the current decision point. Then, the updated baseline price is added to the residual predicted by the model to obtain the predicted price for the current time point. The calculation formula is as follows: ; in, The predicted price output by the model. For the updated baseline price, This is the residual predicted by the model. Subsequently, this predicted price is used as the lagged price input in the feature sequence of the next time point, and the process is advanced point by point to form a near-end prediction sequence with a continuous 15-minute frequency.
[0041] In step S5, a forward macroeconomic base is constructed based on historical real electricity prices and operating characteristics prior to the decision-making point. This base aims to provide a stable price pattern for the forward forecast period, compensating for potential oscillations in the residual LSTM model during long-term extrapolation. The forward macroeconomic base is composed of three parts: a calendar fractal template, an exponentially smoothed extreme value extrapolation curve, and a similar day price curve.
[0042] First, a calendar pattern template is constructed. For each calendar type and weekday combination, historical intraday price curves are extracted and normalized to obtain a standard pattern template. Calendar types include weekdays, weekends, Spring Festival holidays, National Day holidays, pre-holiday windows, post-holiday windows, and adjusted workdays. Intraday segments include morning, noon, and evening segments. For each calendar type and weekday combination, all intraday price curves belonging to that category are extracted from historical data. The mean or median is calculated by time point and normalized to eliminate differences in price levels across different periods, resulting in a standard pattern template reflecting the typical intraday pattern of that calendar type.
[0043] Secondly, an exponential smoothing model is used to predict the price extremes of each intraday segment. For the maximum and minimum prices of each intraday segment, the Holt-Winters exponential smoothing model or other alternative seasonal extrapolation models are used to predict the segment extremes for the next few days based on the historical extreme value sequence. The above standard pattern template is then stretched to the predicted extreme value range, that is, the template is vertically scaled according to the segment extremes of the predicted days, to obtain the template reconstruction curve.
[0044] Next, based on the net load or weather signature for the future forecast period, several similar days are retrieved from the historical sample. The similarity calculation comprehensively considers multiple factors such as the daily average net load, the 96-point net load curve, calendar type, irradiance, cloud cover, wind speed, precipitation, and the proportion of renewable energy. K historical days most similar to the forecast date are selected from the historical sample, where K ranges from 3 to 5. The historical price curves of these K similar days are then averaged or weighted averaged over time points to obtain the similarity day curve.
[0045] Finally, the template reconstruction curve and the similar daily curve are fused according to preset weights to obtain the long-term macroeconomic base. The fusion formula is as follows: ; in, For time points The long-term macroeconomic base forecast value, Reconstruct the curve for the template at time points The value, For similar daily curves at time points The value, The template base weights range from 0 to 1 and can be preset based on historical backtesting results, for example, 0.65. This forward macroeconomic base provides a stable and physically plausible price pattern in the forward forecast range and will be input into the next step for dynamic gating fusion with the near-end forecasts of the residual LSTM model.
[0046] In step S6, the proximal prediction sequence of the residual LSTM model obtained in the previous step is... With the long-term macroeconomic foundation Fusion is achieved through dynamic gating. The core of dynamic gating lies in adaptively calculating fusion weights based on the number of forecast days, the proportion of new energy sources, and the intraday range of net load. This makes the near-term interval more reliant on model predictions with higher sensitivity, while the long-term interval relies more on a more stable macroeconomic foundation.
[0047] The dynamic gating weights are determined by three factors: the exponential decay term corresponding to the number of forecast days, the positive gain term corresponding to the proportion of renewable energy, and the positive gain term corresponding to the intraday range of net load. The gating weights are limited to between zero and one. The weight calculation formula is as follows: ; in, For time points The fusion weights for residual LSTM models; To predict the number of days, i.e., from the current decision-making point to the next time point. The number of days spanned; The attenuation coefficient controls the rate at which the weight decreases with the number of predicted days; The proportion of renewable energy is the ratio of the total predicted renewable energy output to the total predicted renewable energy load. The gain coefficient representing the proportion of new energy sources; The intraday range of net load is the difference between the maximum and minimum values of the predicted intraday net load. The attenuation coefficient is the difference in net load range. This means that the calculation result will be restricted to the range [0,1].
[0048] The larger the forecast period, the greater the exponential decay term. The smaller the value, the greater the weight of the long-term macroeconomic base in the long-term forecast; the higher the proportion of new energy or the larger the intraday range of net load, the greater the positive gain term. and The larger the absolute value, the greater the weight of the residual LSTM model in near-term forecasting, because high new energy penetration and large net load fluctuations mean increased market uncertainty, requiring the model to respond more sensitively to recent operational changes.
[0049] After obtaining the dynamic gating weights, the calculation formula for the fusion prediction result is as follows: ; in, The predicted value after dynamic gating fusion. For residual LSTM models at time points The predicted values are obtained. Through this weighted fusion, an automatic balance between short-term sensitivity and long-term stability is achieved.
[0050] Furthermore, to avoid abrupt changes between the true price boundary and the extrapolation starting point, an exponentially decaying stitching term is introduced near the decision point. Let the last true price be... The extrapolation starting point is predicted to be Then, the difference in time decay is superimposed at the first M time points. This allows the curve to smoothly transition from historical real prices to the predicted curve. After dynamic gating fusion and breakpoint stitching, a fused prediction sequence is obtained. This sequence will be input into the next step for meteorological morphology and holiday post-correction.
[0051] In step S7, historical 15-minute meteorological data and future hourly meteorological data are first loaded. The future hourly meteorological data is mapped to the 15-minute grid using a forward hold method, without unnecessary interpolation, to preserve the original release time meaning of the forecast data. Core meteorological variables include temperature, 80-meter wind speed, cloud cover, solar irradiance, and hourly precipitation.
[0052] A historical meteorological price pattern database is constructed based on historical meteorological data and historical real-time prices. For each historical natural day in this database, the average noon solar irradiance, average cloud cover, average nighttime wind speed, total precipitation, average temperature, renewable energy ratio, average net load, average real-time price, maximum value, minimum value, and the intraday price pattern curve after removing the average are calculated. This information collectively constitutes the weather signature and intraday price pattern curve for that historical natural day, and is stored in the pattern database.
[0053] For each predicted natural day, a weather signature is calculated, which includes at least the average noon solar irradiance, average cloud cover, average nighttime wind speed, total precipitation, and average temperature. Several historical weather days with the most similar weather signature are retrieved from a historical meteorological price pattern database. The intraday price pattern curves of these similar weather days are averaged or weighted averaged over time to construct a weather prototype curve. This curve reflects the average intraday price pattern under similar weather conditions.
[0054] Next, based on one or more of the following factors: irradiance intensity, cloud cover intensity, rainfall intensity, nighttime wind intensity, and heat or cooling load intensity, the fused prediction results are analyzed. The intraday price pattern will be subject to directional correction. Directional corrections include: lowering midday prices during periods of strong sunlight, as strong sunlight increases renewable energy output and lowers electricity prices; reducing the midday price decrease during periods of thick clouds or rain, as clouds and rain reduce solar radiation and decrease renewable energy output; adjusting off-peak prices during nighttime strong winds, as strong winds may increase wind power output and affect off-peak electricity prices; and raising evening peak prices during periods of high or low temperatures, as extreme temperatures increase air conditioning or heating loads and push up evening peak electricity prices. The correction amount is denoted as... , which represents the morphological adjustment of meteorological factors to the fused forecast value.
[0055] Then, a holiday weakening factor is introduced. Fusion factor in the prediction phase The holiday attenuation factor has different attenuation coefficients for different types of holidays: the first attenuation coefficient is used during the Spring Festival and National Day holidays; the second attenuation coefficient is used during the pre-holiday and post-holiday windows; the third attenuation coefficient is used during ordinary weekends; and the fourth attenuation coefficient is used during adjusted workdays. The first attenuation coefficient is less than the second, the second is less than the third, and the third is less than the fourth. This attenuation coefficient is multiplied by a meteorological correction to reduce the intensity of the impact of weather on electricity prices during holidays, because electricity consumption patterns and market behavior change during long holidays, and the meteorological transmission patterns of ordinary days no longer apply.
[0056] Prediction phase fusion factor Based on the forecast period settings: smaller values are used for the near-term forecast interval, intermediate values for the transition period, and larger values for the long-term forecast interval. This is because in the near term, the residual LSTM model has already learned meteorological information through the input features and does not require much post-correction; while in the long term, the macroscopic base is not sensitive to weather and requires stronger post-correction to reflect the impact of weather.
[0057] Finally, for scenarios where strong sunlight, a high proportion of renewable energy, and low net load simultaneously meet preset conditions, a negative price release rule is triggered, allowing midday prices to move towards historical negative price percentiles. However, this is simultaneously constrained by the lower limit of market prices to avoid over-extrapolation. The final formula for calculating the revised forecast result is as follows: ; in, This is the final revised forecast price. As the lower limit of market price, This represents the upper limit of the market price. The function limits the results to a price range allowed by the market. Through the aforementioned post-processing morphological corrections, the final prediction curve retains the basic judgments of the model and its foundation while enhancing the peak-valley variations under high-energy weather scenarios. Furthermore, holiday mitigation and negative price release rules ensure the physical rationality and market compliance of the prediction results. This final prediction result will be input into the next step for diagnostic export, historical backtesting, hourly aggregation, and database write-back.
[0058] Experimental Verification: The parameters and methods of the comparative experiment, as well as data preprocessing, feature engineering, and leakage prevention, were exactly the same. The LSTM network was the baseline algorithm. The LSTM network improved by adding a residual structure is a residual LSTM network, abbreviated as Res-LSTM. An extrapolation model similar to Holt-Winters and similar daily templates were added to construct and optimize the residual LSTM network of the previous step, named HWK-Res-LSTM. Finally, a dynamic gating mechanism (named DG) was added to the HWK-Res-LSTM to form the final model of this paper, named HWK-DG-Res-LSTM. The above LSTM network models optimized by different methods were compared and tested under completely consistent training sets, validation sets, data preprocessing, and leakage prevention feature engineering. The experimental results are shown in Table 1.
[0059] Table 1 Comparison of Ablation Experiments
[0060] As shown in Table 1, from an overall trend perspective, the mean absolute error (MAE) of the baseline LSTM model is 99.74, and the coefficient of determination (R²) is 0.5236. However, when a residual structure is introduced alone and the learning objective is changed from absolute value to deviation relative to the baseline (Res-LSTM), the model performance not only fails to improve as expected but also degrades significantly, with the MAE jumping to 120.58 and the R² slightly dropping to 0.5209. Furthermore, when Holt-Winters-based exponential smoothing extreme value extrapolation and similar daily price curves are superimposed on the residual network (HWK-Res-LSTM), the model accuracy is still only barely on par with Res-LSTM, with an MAE of 120.12 and an R² of 0.5235, which hardly provides any effective improvement to the baseline. It is evident that relying solely on Local Time Series Networks (Res-LSTM) is insufficient to capture high-frequency fluctuations driven by extreme weather and sudden supply-demand changes, and unreasonable residual objectives can amplify prediction bias. While introducing a historically based long-term macroeconomic base (HWK) provides physically reasonable long-term pattern constraints, its inherent smooth regression of historical means lacks sensitivity to sudden market changes near decision points. Therefore, neither the HWK nor the HWK-Res-LSTM can provide substantial gains, whether independently or in simple concatenation. However, when a Dynamic Gated Fusion (DG) mechanism is further introduced on top of HWK-Res-LSTM, the model's predictive performance undergoes a qualitative leap. The resulting HWK-DG-Res-LSTM model drastically reduces the MAE to 54.60, a 45.3% decrease compared to the baseline LSTM, while significantly increasing R² to 0.6138, a 17.2% improvement over the baseline. This significant performance inflection point fully demonstrates that the dynamic gating mechanism is not a simple linear weighter, but rather an adaptive function that accurately achieves the synergistic complementarity of a highly sensitive residual LSTM response in the near-term prediction interval and a smooth transition to a robust macroscopic base structure in the far-term prediction interval. This is achieved by constructing adaptive functions for the exponential decay term of the prediction days, the positive gain term of the new energy proportion, and the intraday range modulation term of the net load. Simultaneously, the exponential decay stitching term near the decision point effectively eliminates the historical true price breakpoint at the extrapolation starting point, allowing the two seemingly inefficient independent modules (Res-LSTM and HWK base) to undergo a strong positive chemical reaction under the regulation of dynamic gating, ultimately achieving a deep coupling effect. In summary, the ablation experiment not only eliminates the randomness of module stacking but also powerfully verifies the feasibility of this invention. The residual structure provides the model with a bias-oriented learning paradigm, the long-term macroscopic base provides a stable physical boundary for long-term prediction, and dynamic gating is the component that activates the potential of both.
[0061] like Figure 3As shown in the figure, Date represents the date, Actual RT Price represents the actual real-time electricity price, Predicted RT Price represents the predicted real-time electricity price, and a fixed column of Anchors represents the anchored benchmark price. RT is an abbreviation for Real-Time, indicating the real-time market settlement price; Anchor represents a fixed reference price based on 00:00 on April 20, 2026, which is typically expressed in currency per megawatt-hour in electricity trading. Figure 3 As can be seen from the shape, the prediction curve of the algorithm of this invention closely matches the overall pattern of actual electricity price cancellation, and the prediction of the trough area is completely correct with an error of less than 1%. For all data, except for a few high points with large prediction deviations, the prediction is accurate and captures the overall pattern. This shows that the method of this invention has good prediction performance in the case of rapid changes in electricity price fluctuations.
[0062] Example 2: A nodal real-time electricity price forecasting system based on multi-source time-series fusion and meteorological adaptive correction. This system is used to implement the nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction described in Example 1, such as... Figure 3 As shown, it includes a data acquisition module, a standardization module, a feature construction module, a model training module, a base construction module, a gating fusion module, and a final prediction module.
[0063] The system comprises the following modules: a data acquisition module for acquiring multi-source power data, including at least historical real-time electricity prices, day-ahead prices, total power generation output forecasts, renewable energy output forecasts, load forecasts, historical meteorological data, and future meteorological data; a standardization module for standardizing multi-source power data into a unified time grid and unifying the target columns for real-time electricity prices and the characteristic standards for renewable energy; a feature construction module for constructing a leak-free feature set with the decision-making time point as the boundary, where real-time electricity prices after the decision-making time point and future unseen features are masked, retaining only the feature data published before the decision-making time point; and a model training module for training a residual LSTM model based on the leak-free feature set. The LSTM model uses the residual of real-time electricity price relative to the baseline price as the learning objective. The foundation construction module is used to construct the long-term macro foundation, which includes a calendar fractal template based on historical prices, an exponentially smoothed extreme value extrapolation curve, and a similar day price curve. The gating fusion module is used to dynamically gating and fusing the prediction results of the residual LSTM model with the long-term macro foundation, where the gating weights are adaptively calculated based on the number of prediction days, the proportion of renewable energy, and net load fluctuations. The final prediction module is used to perform post-form-level corrections on the fused prediction results based on the historical meteorological price pattern library and future meteorological data, and combine holiday weakening factors and negative price release rules to obtain the final prediction result.
[0064] Working Principle: This invention acquires multi-source power data and sequentially executes steps including time standardization and caliber unification, visibility isolation at decision points, residual LSTM near-end prediction, long-term macroeconomic base construction, dynamic gating fusion, and post-processing meteorological morphology correction, constructing a complete closed-loop link from data access to final prediction. First, strict visibility isolation eliminates future data leakage, ensuring backtesting and online prediction are in the same information state. Then, a dual-track parallel approach using residual LSTM and the long-term base handles near-end sensitive response and long-term stable morphology respectively. Next, dynamic gating adaptively switches the dominant force based on the number of prediction days, the proportion of renewable energy, and net load fluctuations. Finally, a morphological correction is applied to the fusion results using a meteorological morphology library, holiday weakening factors, and negative price release rules to adapt to specific scenarios. This invention achieves leak-free, near-end sensitive, long-term stable, meteorologically adaptive, and backtestable real-time node electricity price prediction.
[0065] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0066] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction, characterized in that, Includes the following steps: Acquire multi-source power data, which includes at least historical real-time electricity prices, day-ahead prices, total power generation output forecasts, new energy power generation output forecasts, load forecasts, historical meteorological data, and future meteorological data; The multi-source power data is standardized into a unified time grid, and the real-time electricity price target column and the new energy characteristic caliber are unified. Using the decision-making time point as the boundary, a leak-free feature set is constructed, wherein the real-time electricity price after the decision-making time point and the future invisible features are masked, and only the feature data published before the decision-making time point is retained; Based on the leak-free feature set, a residual LSTM model is trained, with the residual of the real-time electricity price relative to the baseline price as the learning target. Construct a forward macroeconomic foundation, which includes a calendar fractal template based on historical prices, an exponential smoothing extreme value extrapolation curve, and a similar day price curve; The prediction results of the residual LSTM model are fused with the long-term macroeconomic base through dynamic gating, wherein the gating weights are adaptively calculated based on the number of prediction days, the proportion of new energy sources, and net load fluctuations. Based on historical meteorological price pattern database and future meteorological data, a post-form-level correction is performed on the fused forecast results, and combined with holiday weakening factors and negative price release rules, the final forecast results are obtained.
2. The nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction according to claim 1, characterized in that, The construction of the leak-free feature set includes: The decision-making time point simultaneously constrains the available data boundaries of the training set, calendar classification template library, similar day historical sample library, and historical meteorological price pattern library; The actual real-time electricity price after the decision point is only used as a backtesting label or post-evaluation indicator and does not enter any training, template construction, similar day retrieval or morphology library calibration process. The future invisible features include day-ahead prices and daily forecasts that have not yet been published after the decision-making time.
3. The nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction according to claim 1, characterized in that, The definition of new energy characteristics is standardized, including: The effective total new energy volume is selected in descending order of business priority, namely, the sum of the new energy sub-type columns, the total new energy volume column, the daily new energy forecast column, and the weekly new energy forecast column. Furthermore, low-priority data will not be accumulated when high-priority data is available.
4. The nodal real-time electricity price forecasting method based on multi-source time series fusion and meteorological adaptive correction according to claim 1, characterized in that, The baseline price is based on the historical date price or business benchmark price during the training phase, and is dynamically updated based on the latest visible real-time electricity price during the rolling inference phase. The feature sequence input to the residual LSTM model includes at least the hourly index, weekday index, net load, renewable energy percentage, currently used baseline price, lagged price, and rolling statistics. After training is completed, rolling inference is performed on the prediction interval. During each inference, the baseline price is updated based on the latest real-time electricity price published before the current decision point. The updated baseline price is added to the residual predicted by the model to obtain the predicted price at the current time point. This predicted price is then used as the lagged price input in the feature sequence at the next time point, and the process is repeated point by point to form a continuous 15-minute near-end prediction sequence.
5. The nodal real-time electricity price forecasting method based on multi-source time series fusion and meteorological adaptive correction according to claim 1, characterized in that, The construction of the long-term macroeconomic foundation includes: For each calendar type and weekday combination, extract the historical intraday price curve and normalize it to obtain a standard pattern template. An exponential smoothing model is used to predict the price extremes of each intraday segment, and the standard pattern template is stretched to the predicted extreme value range to obtain the template reconstruction curve; Based on the net load or weather signature of the future forecast period, retrieve several similar days from the historical sample, and calculate the average or weighted average of the historical price curves of the similar days to obtain the similar day curves; The template reconstruction curve and the similar day curve are fused together according to a preset weight to obtain the long-term macroeconomic base.
6. The nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction according to claim 1, characterized in that, The weights of the dynamic gating are calculated as follows: The gating weight is jointly determined by the exponential decay term corresponding to the number of forecast days, the positive gain term corresponding to the proportion of new energy, and the positive gain term corresponding to the intraday range of net load, and the gating weight is restricted to the range [0,1]. The larger the number of days to forecast, the smaller the exponential decay term, which increases the weight of the long-term macroeconomic base in the long-term forecast. The higher the proportion of new energy sources or the larger the intraday range of net load, the larger the positive gain term, which increases the weight of the residual LSTM model in near-end prediction. The fusion prediction result is a weighted sum of the prediction results of the residual LSTM model and the long-term macroeconomic base.
7. The nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction according to claim 1, characterized in that, The post-morphological level correction includes: A historical meteorological price pattern database is constructed based on historical meteorological data and historical real-time prices. Each historical natural day in the database contains a weather signature and an intraday price pattern curve. The weather signature includes at least the average noon solar irradiance, average cloud cover, average nighttime wind speed, total precipitation, and average temperature. For a predicted natural day, its weather signature is calculated, similar weather days are retrieved from the morphology library, and a weather prototype curve is constructed. Based on one or more of the following factors: irradiance, cloud cover, rainfall, nighttime wind intensity, and heat or cold load intensity, directional corrections are applied to the intraday price pattern of the fused forecast results. These directional corrections include lowering midday prices during periods of strong sunshine, reducing the midday price drop during periods of thick clouds or rainfall, adjusting trough prices during periods of strong nighttime winds, and raising evening peak prices during periods of high or low temperatures.
8. The nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction according to claim 1, characterized in that, The holiday weakening factor takes a first weakening coefficient during the Spring Festival and National Day holidays, a second weakening coefficient during the pre-holiday and post-holiday windows, a third weakening coefficient during ordinary weekends, and a fourth weakening coefficient during adjusted workdays. The first weakening coefficient is less than the second weakening coefficient, the second weakening coefficient is less than the third weakening coefficient, and the third weakening coefficient is less than the fourth weakening coefficient. The negative price release rule is triggered when strong sunlight, a high proportion of new energy sources, and low net load all meet preset conditions simultaneously, causing the price during the midday period to be released towards the historical negative price percentile, and is constrained by the lower limit of the market price.
9. The nodal real-time electricity price forecasting method based on multi-source time-series fusion and meteorological adaptive correction according to claim 1, characterized in that, The standardization of the multi-source power data into a unified time grid includes: The date field and the time field are combined into a timestamp index. When the value of the time field is 24:00, it is converted to the next day's midnight as the modeling index, and in the final output, it is restored to the 24:00 record of the previous trading day. The final prediction results are aggregated by hourly right-closed intervals and written to the database. Before writing, old prediction records within the target date range are deleted.
10. A nodal real-time electricity price forecasting system based on multi-source time-series fusion and meteorological adaptive correction, characterized in that, include: The data acquisition module is used to acquire multi-source power data, which includes at least historical real-time electricity prices, day-ahead prices, total power generation output forecasts, new energy output forecasts, load forecasts, historical meteorological data, and future meteorological data. The unification module is used to standardize the multi-source power data into a unified time grid and unify the real-time electricity price target column and the new energy characteristic caliber. The feature construction module is used to construct a leak-free feature set with the decision time point as the boundary. The real-time electricity price after the decision time point and the future invisible features are masked, and only the feature data published before the decision time point is retained. The model training module is used to train a residual LSTM model based on the leak-free feature set, wherein the residual LSTM model uses the residual of the real-time electricity price relative to the baseline price as the learning target. The base construction module is used to construct a forward macroeconomic base, which includes a calendar fractal template based on historical prices, an exponential smoothing extreme value extrapolation curve, and a similar day price curve. The gating fusion module is used to dynamically gating and fusing the prediction results of the residual LSTM model with the long-term macroeconomic base, wherein the gating weights are adaptively calculated based on the number of prediction days, the proportion of new energy sources, and net load fluctuations. The final prediction module is used to perform post-form-level corrections on the fused prediction results based on the historical meteorological price pattern database and future meteorological data, and to obtain the final prediction results by combining the holiday weakening factor and the negative price release rule.
