A day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating

Abstract

This invention discloses a day-ahead electricity price forecasting method based on multi-dimensional adaptive gating of operating conditions. The method includes: performing dynamic interpolation repair preprocessing on a day-ahead time-series dataset containing original electricity prices and power-related elements; calculating the global macro-trend comprehensive similarity using weighted grey relational degree with maximum mutual information coefficient; simultaneously constructing a multi-dimensional operating condition fingerprint space and using an asymmetric operating condition adaptive gating mechanism to solve the group confidence distance in real time; when a clustered operating condition threshold is met, gating is activated and nonlinear weight compensation is performed to penalize heterogeneous operating conditions spatially; when facing unknown operating conditions, the gating mechanism adaptively performs zero-axis collapse, allowing the system to smoothly degenerate to the macro-trend relational space; the gating-weighted feature manifold is input into a deep multi-scale time-series coupled mapping network to output the day-ahead electricity price forecast curve. This invention significantly improves the accuracy of electricity price forecasting under specific operating conditions and can provide data support for the stable operation of the power grid.

Description

Technical Field

[0001] This invention belongs to the field of power system operation control and power system dispatch automation technology, and in particular, it is a day-ahead electricity price prediction method based on multi-dimensional operating condition adaptive gating. Background Technology

[0002] The coordinated control and operation optimization of new power systems is a core direction for the current development of the power industry. Day-ahead electricity price time-series data serves as a key control parameter and core information carrier for day-ahead source-grid-load-storage dispatching decisions, rolling optimization of grid operation status, and management of system safety margins in new power systems. Its prediction accuracy directly affects the grid company's safe dispatching efficiency, the coordinated and stable operation of multi-energy complementary systems, and the efficient dynamic allocation of power resources. It plays a crucial supporting role in ensuring the safe and stable operation of the power system under large-scale renewable energy consumption.

[0003] With the large-scale and high-proportion grid connection of renewable energy, the diversification of electricity load demand, and its dramatic nonlinear fluctuations, day-ahead electricity price time series exhibits extremely significant non-stationarity, strong abrupt changes, and heterogeneous spatial coupling characteristics. Its evolution includes strong temporal periodicity caused by the system's inherent operating boundaries and cyclical electricity consumption habits, as well as abrupt changes and fluctuations caused by complex physical conditions such as random flicker in renewable energy output, sudden large load ramp-ups, and alterations to blocked tie lines. Furthermore, it has a deep spatial manifold coupling with key characteristics of multi-dimensional source-grid-load-storage operation, including day-ahead direct-dispatch load, wind, solar, and nuclear power generation, output of non-market-based units, and day-ahead meteorological factors. These complex characteristics pose significant technical challenges to traditional day-ahead electricity price forecasting methods.

[0004] To improve the accuracy of day-ahead electricity price forecasts, research has been conducted in multiple directions. Traditional time-series statistical methods, based on linear stationary laws, are significantly inadequate in fitting electricity price sequences with strong nonlinearity and abrupt changes. Single neural networks or conventional machine learning methods, relying on the generalization ability of deep learning, can capture some time-series dependencies, but when directly inputting the entire set of unordered historical operating data, they are prone to feature learning redundancy and cannot dynamically focus on high-value prior operating conditions.

[0005] In recent years, predictive techniques that incorporate historical similar operating conditions or similar days have become a research hotspot. However, existing technologies have revealed several deep-seated technical deficiencies in practical applications. The quantification of the contributions of multi-dimensional operating factors is often inaccurate. Most existing similar operating condition matching schemes use equal weighting or subjective weighting methods to calculate feature geometric distances, failing to accurately quantify and differentiate the nonlinear differential contributions of different power operating factors to day-ahead electricity price formation under different fluctuation scenarios, resulting in insufficient matching degree of the selected similar operating condition samples. Gating switching and defense mechanisms under abrupt heterogeneous operating conditions are completely lacking in existing research. Existing screening mechanisms are rigid, hard-coded filters, lacking a global judgment of the confidence level of the current day's operating conditions within the historical distribution of the overall market. When faced with highly clustered homogeneous samples in the historical database, the model cannot adaptively amplify the penalty gain of micro-operating conditions. Furthermore, when facing extreme abrupt changes in the system or specific heterogeneous operating conditions, the model is prone to blindly relying on historical paths, forcibly matching unrelated similar days, leading to severe fitting distortion or even prediction collapse in subsequent deep learning networks. There is a lack of robust fallback and adaptive zero-axis collapse defense mechanisms. Furthermore, the incompatibility of the cross-temporal high-dimensional feature coupling modeling mechanism also limits the performance of the model. Existing methods often split similar days into discrete samples after extraction, lacking a unified spatial tensor stacking structure of cross-temporal historical running features and current-day prediction features. This makes it difficult for the subsequent back-end fitting network to efficiently and comprehensively extract multi-scale temporal morphology and long-term, short-term, and long-term time series evolution rules, resulting in limited generalization ability and response speed of the prediction system.

[0006] Therefore, there is an urgent need to develop a day-ahead electricity price forecasting method and system that can accurately screen high-matching historical operating conditions, fully quantify the nonlinear contribution of core driving factors of electricity prices, have adaptive gating conversion capability for heterogeneous and sudden operating conditions, and have strong high-dimensional feature time series capture capability, so as to solve the problems of prediction distortion, weak adaptability to special operating conditions, and poor generalization and defense capabilities of existing technologies. Summary of the Invention

[0007] The purpose of this invention is to address the deficiencies or shortcomings of the existing technology by providing a day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating.

[0008] The technical solution to achieve the objective of this invention is: a day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating, the method comprising the following steps:

[0009] Step 1, Time Series Repair and Preprocessing of Multidimensional Operating Condition Dataset: Collect historical daytime electricity price time series data and multidimensional power operation related element data, perform time series interpolation and repair of local missing values ​​on the original multidimensional time series data, eliminate time series fault interference, and obtain a benchmark full-sample dataset for constructing operating condition fingerprint;

[0010] Step 2, Weighted and Manifold Construction of Operating Condition Features Based on Adaptive Gating: An objective weighting algorithm is used to calculate the nonlinear correlation contribution between each operating condition auxiliary feature and the electricity price sequence in the multidimensional power operation related element data, and the nonlinear correlation contribution is mapped to the dynamic feature weights of each operating condition auxiliary feature; the grey relational analysis method is used to calculate the local correlation coefficients of the fingerprint vectors of historical candidate day operating conditions and the operating condition of the test day in each dimension, and the dynamic feature weights are combined to perform a weighted summation of each local correlation coefficient to obtain the global macro-trend correlation degree; the operating condition fingerprint vectors of the test day and historical candidate days are constructed, and the group confidence distance between the operating condition fingerprint vector of the test day and the operating conditions of historical candidate days in the feature space is calculated. The group confidence distance is thresholded using an asymmetric operating condition adaptive gating mechanism to determine the final similarity; the final similarity is sorted in descending order, and the Top-K historical dates are extracted as target operating condition similarity day samples. The corresponding time series features are spatially stacked to construct a high-dimensional time series feature manifold matrix;

[0011] Step 3, Construction of Deep Multi-Scale Temporal Coupled Mapping Network: Construct a deep multi-scale temporal coupled mapping network, which includes a dilated convolutional module for capturing local multi-scale temporal morphology, a dual-branch network structure for capturing temporal bidirectional dependencies, and an attention mechanism for deep feature cross-fusion.

[0012] Step 4, Continuous mapping inference and day-ahead electricity price output: Input the high-dimensional time-series feature manifold matrix constructed in Step 2 into the trained deep multi-scale time-series coupled mapping network. Through the network, the operating condition-specific prior laws and multi-dimensional spatial coupling relationships are implicitly extracted in the mapping space. Time-series autoregressive evolution decoding is performed to output the day-ahead continuous hourly electricity price prediction curves and obtain the final day-ahead electricity price prediction value.

[0013] Furthermore, step 1 specifically includes:

[0014] Step 1-1: Collaboratively collect the historical day-ahead electricity price time-series data, and simultaneously collect multi-dimensional power operation related element data that are strongly correlated with the day-ahead electricity price; the multi-dimensional power operation related element data includes: day-ahead supply and demand time-series data, real-time key operation data, and day-ahead meteorological time-series data;

[0015] Steps 1-2 involve performing a full-dimensional integrity retrieval on the collected multidimensional time-series data to accurately locate missing time-series positions and statistically analyze the span of time-series faults. A multidimensional time-series dynamic interpolation algorithm is then used to construct an interpolation benchmark based on historically known time-series feature points before and after the missing positions. Smooth interpolation is then performed on all missing values ​​in the multidimensional power operation related element data and electricity price time series to complete the data, outputting a standard electricity price time series and a multidimensional operating condition auxiliary feature dataset.

[0016] Furthermore, the day-ahead supply and demand time-series data includes: day-ahead direct dispatch load, day-ahead tie-line receiving load, day-ahead wind power, day-ahead photovoltaic power, day-ahead nuclear power, and day-ahead non-market-based unit power generation data; the real-time key operation data includes: real-time direct dispatch load, real-time tie-line receiving load, real-time wind power, real-time photovoltaic power, real-time nuclear power, and real-time non-market-based unit power generation data; the day-ahead meteorological time-series data includes: real-time temperature, real-time solar irradiance, and real-time wind speed.

[0017] Furthermore, in step 2, the nonlinear correlation contribution is calculated using an objective weighting algorithm, specifically as follows:

[0018] The maximum mutual information coefficient is used to calculate the time series of the i-th auxiliary feature in the benchmark full sample dataset. Maximum mutual information coefficient between the current day electricity price series y and the current day electricity price series y :

[0019]

[0020] In the formula, For characteristic variables Mutual information with the electricity price variable y; and These represent the number of rows and columns into which the variable is divided into grids, respectively; B is the upper limit of the grid division.

[0021] The obtained Normalization is performed to obtain the dynamic feature weights of the i-th auxiliary feature. :

[0022]

[0023] In the formula, m is the total number of auxiliary features.

[0024] Furthermore, in step 2, the grey relational analysis method is used to calculate the global macro-trend correlation degree. The calculation formula is as follows:

[0025]

[0026] In the formula, The overall similarity of the global macro trend between the date to be tested and the k-th historical candidate date. Let i be the dynamic feature weights of the auxiliary features. The grey relational coefficient between the k-th candidate day and the i-th auxiliary feature is calculated as follows:

[0027]

[0028] In the formula, It is the minimum difference between two levels. Let be the value of the i-th auxiliary feature for the day to be tested. Let be the value of the k-th historical candidate day on the i-th auxiliary feature; The maximum difference between the two levels; The resolution coefficient.

[0029] Furthermore, in step 2, the group confidence distance is thresholded using an asymmetric adaptive gating mechanism to determine the final similarity, specifically including the following steps:

[0030] Step 2-1: Based on the available predictable sequences, extract statistical features to construct the working condition fingerprint vector for the day to be tested and the historical candidate day k working condition fingerprint vectors; after standardizing the working condition fingerprint vectors, calculate the spatial Euclidean distance between the standardized working condition fingerprint of the day to be tested and the working condition fingerprints of each historical candidate day. The operating condition similarity weights for the k-th historical candidate day are calculated using an exponential kernel function. ;

[0031] Step 2-2: Calculate the spatial Euclidean distance between all historical candidate days and the day to be tested, and extract the set confidence quantile boundary values. ;

[0032] Steps 2-3 introduce the global safety condition maximum threshold boundary. and adaptive gated switch variables Compare quantile boundary values Boundary to the maximum threshold of global safety conditions :

[0033] If the relation is satisfied If the current working condition under test has sufficient clustering determinism in historical space, then the adaptive gating switch variable will be determined. Set to 1;

[0034] If the relation is satisfied If the current test condition is determined to be a historically rare and unique scenario, the adaptive zero-axis collapse of the gating mechanism is triggered, and the adaptive gating switch variable is changed. Set to 0;

[0035] Steps 2-4: Correlation of global macro trends Weight of similar working conditions and adaptive gating switch variables Perform global mathematical composition to construct the final similarity score. :

[0036] .

[0037] Further, in step 2-1, the statistical features of the operating condition fingerprint vector include one or more combinations of the following: predicted load peak / valley / average, predicted load maximum ramp rate, predicted wind and solar power output average / fluctuation, wind and solar power penetration rate, and predicted temperature extreme values.

[0038] Furthermore, step 3 specifically includes:

[0039] Step 3-1: Establish a multi-scale parallel convolutional feature extraction module for the historical feature encoding branch. By setting multiple parallel convolutional paths with different kernel sizes and different dilation rates, the module extracts and concatenates the input historical information feature matrix to output a fused feature vector.

[0040] Step 3-2: Establish a BiLSTM bidirectional temporal modeling module for the historical feature encoding branch, receive the fused feature vector, capture the temporal bidirectional dependency through forward and backward LSTM operation units, and output the splicing result of the bidirectional hidden layer state;

[0041] Step 3-3: Establish the Attention filtering module for the historical feature encoding branch, and use the query, key, and value linear matrix to score and normalize the bidirectional temporal hidden layer state, and finally generate the historical feature context vector.

[0042] Steps 3-4: Establish a BiLSTM bidirectional temporal modeling module for the predictive feature encoding branch, receive the predictive information feature matrix, extract the temporal evolution pattern of the daily predicted data sequence through forward and backward LSTM units, and output the predictive hidden layer state vector.

[0043] Steps 3-5: Establish a feature splicing and fusion layer, splicing and combining the historical feature context vector and the predicted hidden layer state vector in the feature dimension to generate a joint high-dimensional feature vector;

[0044] Steps 3-6: Establish the LSTM self-looping module of the decoder, perform autoregressive decoding on the joint high-dimensional features, and output the hidden layer state of the decoder.

[0045] Steps 3-7 establish the reconstruction prediction module of the decoder. After performing random deactivation processing on the hidden layer state of the decoder, the predicted electricity price values ​​for consecutive hours before the day are output through the output linear layer mapping.

[0046] Furthermore, step 3 also includes configuring the deep multi-scale temporally coupled mapping network: setting the continuous optimization hyperparameters of the deep multi-scale temporally coupled mapping network, using the mean absolute error as the loss function, and configuring a learning rate decay strategy using an adaptive moment estimation optimizer; wherein, the loss function and the learning rate decay mechanism are respectively expressed as:

[0047]

[0048]

[0049] In the formula, N is the total number of predicted samples; This represents the actual electricity price at time t. Here is the predicted electricity price at time t; MAE is the mean absolute error. Let be the learning rate for the t-th training round; The initial learning rate; This is the learning rate decay factor; This is the decay step size; This is the floor function.

[0050] Furthermore, in step 4, the feature data of the day to be measured is input into the trained deep multi-scale temporal coupled mapping network model for prediction to obtain the final electricity price prediction result:

[0051]

[0052] In the formula, The historical information feature matrix of the day before the test date is used as the input to the historical feature encoding branch in the trained deep multi-scale temporal coupled mapping network model; The predicted information feature matrix for the day to be tested is used as the input for the predicted feature encoding branch; This represents a trained deep multi-scale temporally coupled mapping network model based on dual-branch feature fusion and encoder-decoder architecture. This is the predicted electricity price for the 24 hours prior to the test date, directly output by the model after dual-branch feature extraction, data splicing and fusion, and decoding mapping.

[0053] Compared with the prior art, the significant advantages of this invention are:

[0054] (1) This invention breaks through the limitations of traditional similar operating condition matching that uses equal weights or subjective weighting, and adopts an objective weighting screening scheme based on the weighted grey relational degree of the maximum mutual information coefficient. This scheme quantifies the nonlinear differential contribution of different operating elements to the formation of day-ahead electricity prices under the time-of-day division of the whole day by using the maximum mutual information coefficient. Combined with the grey relational analysis method, it calculates the local grey relational coefficients and macro-global relational characteristics of historical candidate days and the day to be tested, and screens out a set of similar days that highly match the day to be tested. This scheme effectively filters out noise interference in historical operating data, significantly reduces the uncertainty of day-ahead electricity price forecasting, and improves the correlation between input features and forecast targets.

[0055] (2) To address historically rare and unique heterogeneous abrupt changes in operating conditions such as extreme ramp-ups and instantaneous shifts in new energy sources, this invention introduces a post-adaptive gating and zero-axis collapse defense conversion scheme based on a multi-dimensional operating condition fingerprint space. This scheme uses a global mathematical composite formula containing gating switch variables to determine the group confidence distance of the current day's overall operating conditions in the historical manifold distribution: under homogeneous clustering conditions, the gating switch variables are activated and amplify the spatial penalty gain of micro-specific operating conditions; under extreme ramp-ups or instantaneous shifts in new energy sources, the gating switch variables trigger adaptive zero-axis collapse and are set to zero, allowing the system to smoothly retreat to the macro-trend similarity space. This eliminates the fitting distortion and prediction collapse caused by the forced matching of irrelevant historical samples under extreme operating conditions, ensuring the elastic noise resistance and stable fallback performance of the predicted overall market.

[0056] (3) This invention abandons the traditional approach of splitting the similar days obtained by screening into discrete samples. Instead, it adopts an integrated cross-temporal tensor stacking scheme based on a high-dimensional temporal feature manifold matrix. The historical domain feature matrix and the prediction domain feature matrix are aligned and stacked across the temporal domain in the spatial dimension to directly construct a high-dimensional temporal feature manifold matrix, which provides a highly adaptable topological input structure for the backend deep learning network.

[0057] (4) This invention constructs a unified deep multi-scale temporal coupling mapping network. Through the dilated convolution module, dual-branch network topology and cross-attention mechanism, it extracts multi-scale temporal morphology, bidirectional long and short-term temporal dependencies and spatial coupling relationships between multi-dimensional elements in a one-stop manner within the unified mapping space. This ensures the fitting accuracy and convergence stability of the daily continuous 24-hour electricity price forecast curve, and provides data support for the rolling decision-making, safety margin configuration and coordinated stable operation of the power system dispatch automation system and the multi-energy complementary system of source, grid, load and storage.

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0059] Figure 1 This is a flowchart of the day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to the present invention.

[0060] Figure 2 This is a detailed flowchart of a similar day screening process based on the working condition fingerprint space and adaptive gating mechanism in one embodiment.

[0061] Figure 3 This is a block diagram of the overall data flow architecture of the system jointly predicted by adaptive gating similar day screening and deep learning network in one embodiment.

[0062] Figure 4 This is a curve comparing the predicted and actual values ​​of the test sample 24 hours prior to the test date in one embodiment. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] It should be noted that if the embodiments of the present invention involve descriptions such as "first" and "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0065] In one embodiment, combined Figures 1 to 3 This paper provides a day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating, the method comprising the following steps:

[0066] Step 1, Time Series Repair and Preprocessing of Multidimensional Operating Condition Dataset: Collect historical daytime electricity price time series data and multidimensional power operation related element data, perform time series interpolation and repair of local missing values ​​on the original multidimensional time series data, eliminate time series fault interference, and obtain a benchmark full-sample dataset for constructing operating condition fingerprint;

[0067] Step 2, Weighted and Manifold Construction of Operating Condition Features Based on Adaptive Gating: An objective weighting algorithm is used to calculate the nonlinear correlation contribution between each operating condition auxiliary feature and the electricity price sequence in the multidimensional power operation related element data, and the nonlinear correlation contribution is mapped to the dynamic feature weights of each operating condition auxiliary feature; the grey relational analysis method is used to calculate the local correlation coefficients of the fingerprint vectors of historical candidate day operating conditions and the operating condition of the test day in each dimension, and the dynamic feature weights are combined to perform a weighted summation of each local correlation coefficient to obtain the global macro-trend correlation degree; the operating condition fingerprint vectors of the test day and historical candidate days are constructed, and the group confidence distance between the operating condition fingerprint vector of the test day and the operating conditions of historical candidate days in the feature space is calculated. The group confidence distance is thresholded using an asymmetric operating condition adaptive gating mechanism to determine the final similarity; the final similarity is sorted in descending order, and the Top-K historical dates are extracted as target operating condition similarity day samples. The corresponding time series features are spatially stacked to construct a high-dimensional time series feature manifold matrix;

[0068] Step 3, Construction of Deep Multi-Scale Temporal Coupled Mapping Network: Construct a deep multi-scale temporal coupled mapping network, which includes a dilated convolutional module for capturing local multi-scale temporal morphology, a dual-branch network structure for capturing temporal bidirectional dependencies, and an attention mechanism for deep feature cross-fusion.

[0069] Step 4, Continuous mapping inference and day-ahead electricity price output: Input the high-dimensional time-series feature manifold matrix constructed in Step 2 into the trained deep multi-scale time-series coupled mapping network. Through the network, the operating condition-specific prior laws and multi-dimensional spatial coupling relationships are implicitly extracted in the mapping space. Time-series autoregressive evolution decoding is performed to output the day-ahead continuous hourly electricity price prediction curves and obtain the final day-ahead electricity price prediction value.

[0070] Furthermore, in one embodiment, step 1 specifically includes:

[0071] Step 1-1: Collaboratively collect the historical day-ahead electricity price time-series data, and simultaneously collect multi-dimensional power operation related element data that are strongly correlated with the day-ahead electricity price; the multi-dimensional power operation related element data includes: day-ahead supply and demand time-series data, real-time key operation data, and day-ahead meteorological time-series data;

[0072] Steps 1-2 involve performing a full-dimensional integrity retrieval on the collected multidimensional time-series data to accurately locate missing time-series positions and statistically analyze the span of time-series faults. A multidimensional time-series dynamic interpolation algorithm is then used to construct an interpolation benchmark based on historically known time-series feature points before and after the missing positions. Smooth interpolation is then performed on all missing values ​​in the multidimensional power operation related element data and electricity price time series to complete the data, outputting a standard electricity price time series and a multidimensional operating condition auxiliary feature dataset.

[0073] Preferably, in some embodiments, the day-ahead supply and demand time-series data includes: day-ahead direct-suppression load, day-ahead tie-line power receiving load, day-ahead wind power, day-ahead photovoltaic power, day-ahead nuclear power, and day-ahead non-market-based unit power generation data; the real-time operational key data includes: real-time direct-suppression load, real-time tie-line power receiving load, real-time wind power, real-time photovoltaic power, real-time nuclear power, and real-time non-market-based unit power generation data; the day-ahead meteorological time-series data includes: real-time temperature, real-time solar irradiance, and real-time wind speed.

[0074] Furthermore, in one embodiment, in step 2, the nonlinear correlation contribution is calculated using an objective weighting algorithm, specifically as follows:

[0075] Feature vectors for the target day and historical candidate days are pre-constructed. For the target day d, the predicted power generation and load values ​​for that day are extracted, and combined with the historical actual power generation, actual load, and actual electricity price of the previous day d−1, to form the feature vector for the target day. Where m is the total number of features; similarly, extract the same-dimensional features of each candidate day k from the historical dataset to construct the feature vector of the historical candidate days. ;

[0076] The maximum mutual information coefficient is used to calculate the time series of the i-th auxiliary feature in the benchmark full sample dataset. Maximum mutual information coefficient between the current day electricity price series y and the current day electricity price series y :

[0077]

[0078] In the formula, For characteristic variables Mutual information with the electricity price variable y; and These represent the number of rows and columns into which the variable is divided into grids, respectively; B is the upper limit of the grid division.

[0079] The obtained Normalization is performed to obtain the dynamic feature weights of the i-th auxiliary feature. :

[0080]

[0081] In the formula, m is the total number of auxiliary features.

[0082] Furthermore, in one embodiment, in step 2, the global macro-trend correlation degree is calculated using the grey relational analysis method, and the calculation formula is as follows:

[0083]

[0084] In the formula, The overall similarity of the global macro trend between the date to be tested and the k-th historical candidate date. Let i be the dynamic feature weights of the auxiliary features. Let be the grey relational coefficient between the k-th candidate day and the i-th auxiliary feature;

[0085] Using the feature vector of the day to be measured As a reference sequence, the eigenvectors of historical candidate days As a comparison sequence, the grey relational coefficient between the k-th candidate day and the test day on the i-th feature is calculated. for:

[0086]

[0087] In the formula, It is the minimum difference between two levels. Let be the value of the i-th auxiliary feature for the day to be tested. Let be the value of the k-th historical candidate day on the i-th auxiliary feature; The maximum difference between the two levels; The resolution coefficient.

[0088] Furthermore, in one embodiment, step 2 involves using an asymmetric adaptive gating mechanism to threshold the group confidence distance to determine the final similarity, specifically including the following steps:

[0089] Step 2-1: Extract statistical features based on the available predictable sequences to construct the daily working condition fingerprint vector to be tested. Where q is the number of working condition fingerprint features) and the historical candidate day k working condition fingerprint vector ( After standardizing the operating condition fingerprint vector, the spatial Euclidean distance between the standardized operating condition fingerprint of the test day and the operating condition fingerprints of each historical candidate day is calculated. The operating condition similarity weights for the k-th historical candidate day are calculated using an exponential kernel function. ;

[0090] Specifically: collect all historical candidate day operation condition fingerprint vectors to form a historical fingerprint matrix, and calculate the full-dimensional statistical mean. and standard deviation The mean and standard deviation are used to analyze the fingerprint vector of the daily operating conditions to be measured. fingerprint vectors of various historical candidate days Implementing standardization:

[0091] ,

[0092] In the formula, To prevent zero smoothing constant;

[0093] Calculate the spatial Euclidean distance between the standardized operating condition fingerprint of the test day and the operating condition fingerprints of each historical candidate day. The operating condition similarity weights for the k-th historical candidate day are calculated using an exponential kernel function. :

[0094] ,

[0095] In the formula, T is the temperature control adjustment parameter, which is set to 1;

[0096] Step 2-2: Calculate the spatial Euclidean distance between all historical candidate days and the day to be tested, and extract the set confidence quantile boundary values. ;

[0097] here, ,in The confidence quantile parameter is preferably set within the range of 5% to 20%.

[0098] Steps 2-3 introduce the global safety condition maximum threshold boundary. and adaptive gated switch variables Compare quantile boundary values Boundary to the maximum threshold of global safety conditions :

[0099] If the relation is satisfied If the current working condition under test has sufficient clustering determinism in historical space, then the adaptive gating switch variable will be determined. Set to 1 to activate the feature penalty correction state;

[0100] If the relation is satisfied If the current test condition is determined to be a historically rare and unique scenario, the adaptive zero-axis collapse of the gating mechanism is triggered, and the adaptive gating switch variable is changed. Set to 0;

[0101] Here, the maximum threshold boundary of the global safety operating condition is pre-calculated as the preset quantile, such as the median, of the Euclidean distance between the pairwise operating condition fingerprints of all samples in the historical training set, which is used to characterize the typical dispersion of the historical operating conditions.

[0102] Steps 2-4: Correlation of global macro trends Weight of similar working conditions and adaptive gating switch variables Perform global mathematical composition to construct the final similarity score. :

[0103] .

[0104] Preferably, in some embodiments, in step 2-1, the statistical features of the operating condition fingerprint vector include one or more combinations of the following: predicted load peak / valley / average, predicted load maximum ramp rate, predicted wind and solar power output average / fluctuation, wind and solar power penetration rate, and predicted temperature extreme values.

[0105] Furthermore, in one embodiment, step 3 specifically includes:

[0106] Step 3-1: Establish a multi-scale parallel convolutional feature extraction module for the historical feature encoding branch. By setting multiple parallel convolutional paths with different kernel sizes and dilation rates, features are extracted from the input historical information feature matrix and concatenated to output a fused feature vector; represented as:

[0107]

[0108] In the formula, The input is a historical information feature matrix; This represents the i-th parallel convolution path; Let be the kernel size of the i-th parallel convolution path; Let be the dilation rate of the i-th parallel convolution path; The intermediate feature representation extracted from the i-th parallel branch; It is a non-linear activation function; This is a multi-branch feature concatenation operation; This is a batch normalization operation used to accelerate model training and improve stability; This is a random deactivation operation used to suppress model overfitting; This is the final fused feature vector output by the module.

[0109] Step 3-2: Establish a BiLSTM bidirectional temporal modeling module for the historical feature encoding branch. This module receives the fused feature vector, captures the temporal bidirectional dependencies through forward and backward LSTM operation units, and outputs the concatenated bidirectional hidden layer states; represented as:

[0110]

[0111] In the formula, Let be the input vector at time t; and These represent the forward and backward LSTM operation units, respectively; and These represent the forward and reverse hidden layer states at time t, respectively. This is the result of splicing the states of the bidirectional hidden layer.

[0112] Step 3-3: Establish the Attention filtering module for the historical feature encoding branch. Utilize the query, key, and value linear matrix to score and normalize the bidirectional temporal hidden layer states, ultimately generating a historical feature context vector; represented as:

[0113]

[0114] In the formula, It is the hidden layer state of the decoder at the current moment; It is the bidirectional temporal hidden layer state output by the historical feature encoding branch; , , , These are the weight matrices for the query, key, value, and output linear layers, respectively. , , , These are the bias terms for the corresponding linear layers; For the hidden layer feature dimension, This is a scaling factor used to avoid gradient anomalies caused by excessively large inner product values; For attention scoring functions; This is a normalization function used to convert the scores into attention weights in the (0,1) interval. ; The context vector obtained by weighted summation of the vector values; This is a random deactivation operation used to suppress model overfitting; The final context vector generated by the attention filtering module is used as the feature input for the subsequent decoder.

[0115] Steps 3-4: Establish a BiLSTM bidirectional temporal modeling module for the predictive feature encoding branch, receive the predictive information feature matrix, extract the temporal evolution pattern of the daily predicted data sequence through forward and backward LSTM units, and output the predictive hidden layer state vector.

[0116] The network structure of this BiLSTM bidirectional temporal modeling module is exactly the same as that of the BiLSTM bidirectional temporal modeling module with the historical feature encoding branch described in step 3-2. It extracts the temporal evolution pattern of the current-day prediction data sequence through forward and backward LSTM units composed of forget gates, input gates, unit states, and output gates. Let its output prediction hidden layer state vector be... ;

[0117] Steps 3-5: Establish a feature concatenation and fusion layer, concatenating the historical feature context vector with the predicted hidden layer state vector along the feature dimension to generate a joint high-dimensional feature vector; represented as:

[0118]

[0119] Steps 3-6: Establish the LSTM self-looping module of the decoder to perform autoregressive decoding on the joint high-dimensional features and output the hidden layer state of the decoder; represented as:

[0120]

[0121] In the formula, This is the input to the decoder at time t; It is the hidden layer output of the decoder at time t-1, the initial time. ; This is the unit state of the decoder at time t-1, the initial time. ; Forgotten Gate; It is the Sigmoid activation function; It is an input gate; This indicates the cell state update value; , , , These are the weight matrices for the forget gate, input gate, cell state update, and output gate, respectively. , , , These are the bias terms for the corresponding gates; tanh is the hyperbolic tangent activation function; It is an output gate; It is the hidden layer output of the decoder at time t, which is used as the input to the subsequent reconstruction prediction module.

[0122] Steps 3-7 establish the decoder's reconstruction prediction module. After performing random deactivation processing on the decoder's hidden layer states, the predicted electricity prices for consecutive hours before the current day are output through the output linear layer mapping. This is represented as:

[0123]

[0124] In the formula, This is the output of the hidden layer of the decoder at time t; This is a random deactivation operation; is the hidden layer state after random deactivation; W is the weight matrix of the output linear layer; b is the bias term of the output linear layer; This is the predicted electricity price at time t, which is the final output of the module.

[0125] Furthermore, in one embodiment, step 3 further includes configuring the deep multi-scale temporally coupled mapping network: setting the continuous optimization hyperparameters of the deep multi-scale temporally coupled mapping network, using the mean absolute error as the loss function, and configuring a learning rate decay strategy using an adaptive moment estimation optimizer; wherein, the loss function and the learning rate decay mechanism are respectively expressed as:

[0126]

[0127]

[0128] In the formula, N is the total number of predicted samples; This represents the actual electricity price at time t. Here is the predicted electricity price at time t; MAE is the mean absolute error. Let be the learning rate for the t-th training round; The initial learning rate; This is the learning rate decay factor; This is the decay step size; This is the floor function.

[0129] Furthermore, in one embodiment, in step 4, the feature data of the day to be measured is input into a trained deep multi-scale temporal coupled mapping network model for prediction to obtain the final electricity price prediction result:

[0130]

[0131] In the formula, The historical information feature matrix of the day before the test date is used as the input to the historical feature encoding branch in the trained deep multi-scale temporal coupled mapping network model; The predicted information feature matrix for the day to be tested is used as the input for the predicted feature encoding branch; This represents a trained deep multi-scale temporally coupled mapping network model based on dual-branch feature fusion and encoder-decoder architecture. This is the predicted electricity price for the 24 hours prior to the test date, directly output by the model after dual-branch feature extraction, data splicing and fusion, and decoding mapping.

[0132] In one embodiment, a day-ahead electricity price forecasting system based on multi-dimensional operating condition adaptive gating is provided, the system comprising:

[0133] The first module is used to realize the time series repair and preprocessing of multi-dimensional operating condition dataset: collect historical day-ahead electricity price time series data and multi-dimensional power operation related element data, perform local missing value time series interpolation and repair on the original multi-dimensional time series data, eliminate time series fault interference, and obtain a benchmark full sample dataset for constructing operating condition fingerprint;

[0134] The second module is used to implement weighted and manifold construction of operating condition features based on adaptive gating: An objective weighting algorithm is used to calculate the nonlinear correlation contribution between each operating condition auxiliary feature and the electricity price sequence in the multidimensional power operation related element data, and the nonlinear correlation contribution is mapped to the dynamic feature weights of each operating condition auxiliary feature; the grey relational analysis method is used to calculate the local correlation coefficients of the fingerprint vectors of historical candidate day operating conditions and the operating condition of the test day in each dimension, and the dynamic feature weights are combined to perform a weighted summation of each local correlation coefficient to obtain the global macro-trend correlation degree; the operating condition fingerprint vectors of the test day and historical candidate days are constructed, and the group confidence distance between the operating condition fingerprint vector of the test day and the operating conditions of historical candidate days in the feature space is calculated. An asymmetric operating condition adaptive gating mechanism is used to threshold the group confidence distance to determine the final similarity; the final similarity is sorted in descending order, and the Top-K historical dates are extracted as target operating condition similarity day samples. The corresponding time series features are spatially stacked to construct a high-dimensional time series feature manifold matrix;

[0135] The third module is used to construct a deep multi-scale temporal coupling mapping network: a deep multi-scale temporal coupling mapping network is constructed, which includes a dilated convolution module for capturing local multi-scale temporal morphology, a dual-branch network structure for capturing temporal bidirectional dependencies, and an attention mechanism for deep feature cross-fusion.

[0136] The fourth module is used to implement continuous mapping inference and day-ahead electricity price output: the high-dimensional time-series feature manifold matrix constructed in the second module is input into the trained deep multi-scale time-series coupled mapping network. The network implicitly extracts the condition-specific prior laws and multi-dimensional spatial coupling relationships in the mapping space, performs time-series autoregressive evolution decoding, outputs the day-ahead continuous hourly electricity price prediction curve, and obtains the final day-ahead electricity price prediction value.

[0137] Specific limitations regarding the day-ahead electricity price forecasting system based on multi-dimensional operating condition adaptive gating can be found in the limitations of the day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating mentioned above, and will not be repeated here. Each module in the above-mentioned day-ahead electricity price forecasting system based on multi-dimensional operating condition adaptive gating can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0138] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0139] Step 1, Time Series Repair and Preprocessing of Multidimensional Operating Condition Dataset: Collect historical daytime electricity price time series data and multidimensional power operation related element data, perform time series interpolation and repair of local missing values ​​on the original multidimensional time series data, eliminate time series fault interference, and obtain a benchmark full-sample dataset for constructing operating condition fingerprint;

[0140] Step 2, Weighted and Manifold Construction of Operating Condition Features Based on Adaptive Gating: An objective weighting algorithm is used to calculate the nonlinear correlation contribution between each operating condition auxiliary feature and the electricity price sequence in the multidimensional power operation related element data, and the nonlinear correlation contribution is mapped to the dynamic feature weights of each operating condition auxiliary feature; the grey relational analysis method is used to calculate the local correlation coefficients of the fingerprint vectors of historical candidate day operating conditions and the operating condition of the test day in each dimension, and the dynamic feature weights are combined to perform a weighted summation of each local correlation coefficient to obtain the global macro-trend correlation degree; the operating condition fingerprint vectors of the test day and historical candidate days are constructed, and the group confidence distance between the operating condition fingerprint vector of the test day and the operating conditions of historical candidate days in the feature space is calculated. The group confidence distance is thresholded using an asymmetric operating condition adaptive gating mechanism to determine the final similarity; the final similarity is sorted in descending order, and the Top-K historical dates are extracted as target operating condition similarity day samples. The corresponding time series features are spatially stacked to construct a high-dimensional time series feature manifold matrix;

[0141] Step 3, Construction of Deep Multi-Scale Temporal Coupled Mapping Network: Construct a deep multi-scale temporal coupled mapping network, which includes a dilated convolutional module for capturing local multi-scale temporal morphology, a dual-branch network structure for capturing temporal bidirectional dependencies, and an attention mechanism for deep feature cross-fusion.

[0142] Step 4, Continuous mapping inference and day-ahead electricity price output: Input the high-dimensional time-series feature manifold matrix constructed in Step 2 into the trained deep multi-scale time-series coupled mapping network. Through the network, the operating condition-specific prior laws and multi-dimensional spatial coupling relationships are implicitly extracted in the mapping space. Time-series autoregressive evolution decoding is performed to output the day-ahead continuous hourly electricity price prediction curves and obtain the final day-ahead electricity price prediction value.

[0143] For specific limitations on each step, please refer to the limitations of the day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating mentioned above, which will not be repeated here.

[0144] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0145] Step 1, Time Series Repair and Preprocessing of Multidimensional Operating Condition Dataset: Collect historical daytime electricity price time series data and multidimensional power operation related element data, perform time series interpolation and repair of local missing values ​​on the original multidimensional time series data, eliminate time series fault interference, and obtain a benchmark full-sample dataset for constructing operating condition fingerprint;

[0146] Step 2, Weighted and Manifold Construction of Operating Condition Features Based on Adaptive Gating: An objective weighting algorithm is used to calculate the nonlinear correlation contribution between each operating condition auxiliary feature and the electricity price sequence in the multidimensional power operation related element data, and the nonlinear correlation contribution is mapped to the dynamic feature weights of each operating condition auxiliary feature; the grey relational analysis method is used to calculate the local correlation coefficients of the fingerprint vectors of historical candidate day operating conditions and the operating condition of the test day in each dimension, and the dynamic feature weights are combined to perform a weighted summation of each local correlation coefficient to obtain the global macro-trend correlation degree; the operating condition fingerprint vectors of the test day and historical candidate days are constructed, and the group confidence distance between the operating condition fingerprint vector of the test day and the operating conditions of historical candidate days in the feature space is calculated. The group confidence distance is thresholded using an asymmetric operating condition adaptive gating mechanism to determine the final similarity; the final similarity is sorted in descending order, and the Top-K historical dates are extracted as target operating condition similarity day samples. The corresponding time series features are spatially stacked to construct a high-dimensional time series feature manifold matrix;

[0147] Step 3, Construction of Deep Multi-Scale Temporal Coupled Mapping Network: Construct a deep multi-scale temporal coupled mapping network, which includes a dilated convolutional module for capturing local multi-scale temporal morphology, a dual-branch network structure for capturing temporal bidirectional dependencies, and an attention mechanism for deep feature cross-fusion.

[0148] Step 4, Continuous mapping inference and day-ahead electricity price output: Input the high-dimensional time-series feature manifold matrix constructed in Step 2 into the trained deep multi-scale time-series coupled mapping network. Through the network, the operating condition-specific prior laws and multi-dimensional spatial coupling relationships are implicitly extracted in the mapping space. Time-series autoregressive evolution decoding is performed to output the day-ahead continuous hourly electricity price prediction curves and obtain the final day-ahead electricity price prediction value.

[0149] For specific limitations on each step, please refer to the limitations of the day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating mentioned above, which will not be repeated here.

[0150] As a specific example, the invention will be further described and verified in detail in one embodiment.

[0151] This embodiment uses time-of-use daytime electricity price time-series data from a specific regional electricity market, along with strongly correlated multidimensional power operation related element data, as the training and testing foundation. The data covers daytime source-grid-load-storage power supply and demand forecasts and historical real data for direct-dispatch load, tie-line load, wind power, photovoltaic power, nuclear power, and non-market-based unit power generation, as well as meteorological time-series data such as temperature, solar irradiance, and wind speed, and date-type feature data. The input sample size is set to the high-dimensional time-series feature manifold matrix constructed from the top 50 historically highly matched similar days, and the prediction target is the electricity price time-series curve for the consecutive 24 hours of the daytime.

[0152] The configuration parameters of the deep multi-scale temporal coupled mapping network model constructed in this embodiment are as follows: the multi-scale parallel feature extraction module consists of 8 parallel one-dimensional convolutional branches, covering standard convolutions with different receptive fields and dilated convolutions with different dilation rates, and the output dimension of a single branch convolution is set to 16; the network hidden layer is configured with 2 layers of bidirectional recurrent computation units, each containing 128 neurons; the input and output dimensions of the cross attention mechanism module are aligned to 128 dimensions to strengthen key temporal morphological weights; the decoder is configured with 2 layers of unidirectional recurrent computation units, each containing 128 neurons.

[0153] During the model training and configuration phase, the evaluation and loss function were set as mean absolute error, and the network optimizer was an adaptive moment estimator. Specifically, the network Dropout regularization rate was set to 0.3, the initial learning rate to 0.002, and a learning rate decay strategy was configured with a decay step size of 10 and a decay factor of 0.8. The batch size during training was set to 32, and the maximum number of training iterations was set to 40 to ensure the stability of the convergence of the multi-scale network topology parameters.

[0154] In this embodiment, the last 5 trading days in the standard time series dataset are used as test samples to compare and verify the performance of the proposed day-ahead electricity price prediction based on multi-dimensional operating condition adaptive gating. After multiple tests and averaging, the prediction error statistics of each method are shown in Table 1.

[0155] Table 1. Statistical results of prediction errors for each method

[0156]

[0157] As shown in Table 1, the test results of this invention demonstrate significant superiority in both prediction accuracy and control robustness for the 24 hours prior to the test sample. Overall, the mean absolute error (MAE) of the full-sample direct training method is 60.09, and the mean squared error (MSE) is 6906.10. The conventional similar day screening method based on MIC-GRA filters out some historical noise through objective feature weighting, reducing its mean MAE and mean MSE to 50.16 and 5541.59, respectively. Furthermore, the method of this invention, which combines condition gating and MIC-GRA, introduces adaptive gating switch variables and a zero-axis collapse defense mechanism, resulting in an overwhelming convergence of the full-sample mean MAE to 46.35 and a simultaneous decrease in the mean MSE to 4796.04.

[0158] In samples 3, 4, and 5 with a gating state of 1, the method of this invention, relying on the spatial penalty gain of microscopic specific operating conditions, achieved a continuous stepwise decrease in various error indicators. Among them, the MAE and MSE of sample 5 were as low as 21.01 and 1391.79, respectively, demonstrating excellent global prediction bias control. In samples 1 and 2, where the system encountered extreme fluctuations or rare operating conditions that caused the gating state to be set to zero, the method of this invention successfully triggered an adaptive zero-direction collapse mechanism and smoothly retreated to the macroscopic trend space, suppressing fitting distortion and large error propagation under specific operating conditions. This fully verifies that the method of this invention has extremely high prediction accuracy and dynamic stability fallback performance under all operating conditions.

[0159] from Figure 4 The curve fitting results show that the predicted values ​​output by the model of this invention are highly consistent with the changing trends of the actual electricity prices. It can accurately reproduce the typical patterns of diversified and drastic intraday fluctuations in day-ahead electricity prices in specific regional power systems. In particular, it achieves high-precision integrated fitting for local negative electricity price troughs caused by high-proportion renewable energy grid connection, rapid rise and fall inflection points triggered by heavy load ramp-up, and peak electricity prices during peak hours. Test results show that this invention effectively reduces the prediction uncertainty caused by the heterogeneous spatial coupling and non-stationary characteristics of day-ahead electricity price sequences. It has a strong dynamic capture capability for electricity price mutations under extreme and specific system operating conditions. The stability and generalization ability of time series prediction are excellent, and it can accurately capture the evolution of the supply and demand patterns of power resources.

[0160] The high-precision and high-stability prediction curves output by this invention can provide reliable control parameter support for the formulation of day-ahead multi-energy complementary rolling dispatch plans, source-grid-load-storage coordinated control decisions, and the refined configuration of grid safety margin and reserve capacity in new power systems, effectively promoting the optimized allocation and safe and stable operation of power resources in new power systems.

[0161] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating, characterized in that, The method includes the following steps: Step 1, Time Series Repair and Preprocessing of Multidimensional Operating Condition Dataset: Collect historical daytime electricity price time series data and multidimensional power operation related element data, perform time series interpolation and repair of local missing values ​​on the original multidimensional time series data, eliminate time series fault interference, and obtain a benchmark full-sample dataset for constructing operating condition fingerprint; Step 2, Weighted and Manifold Construction of Operating Condition Features Based on Adaptive Gating: An objective weighting algorithm is used to calculate the nonlinear correlation contribution between each operating condition auxiliary feature and the electricity price sequence in the multidimensional power operation related element data, and the nonlinear correlation contribution is mapped to the dynamic feature weights of each operating condition auxiliary feature; the grey relational analysis method is used to calculate the local correlation coefficients of the fingerprint vectors of historical candidate day operating conditions and the operating condition of the test day in each dimension, and the dynamic feature weights are combined to perform a weighted summation of each local correlation coefficient to obtain the global macro-trend correlation degree; the operating condition fingerprint vectors of the test day and historical candidate days are constructed, and the group confidence distance between the operating condition fingerprint vector of the test day and the operating conditions of historical candidate days in the feature space is calculated. The group confidence distance is thresholded using an asymmetric operating condition adaptive gating mechanism to determine the final similarity; the final similarity is sorted in descending order, and the Top-K historical dates are extracted as target operating condition similarity day samples. The corresponding time series features are spatially stacked to construct a high-dimensional time series feature manifold matrix; Step 3, Construction of Deep Multi-Scale Temporal Coupled Mapping Network: Construct a deep multi-scale temporal coupled mapping network, which includes a dilated convolutional module for capturing local multi-scale temporal morphology, a dual-branch network structure for capturing temporal bidirectional dependencies, and an attention mechanism for deep feature cross-fusion. Step 4, Continuous mapping inference and day-ahead electricity price output: Input the high-dimensional time-series feature manifold matrix constructed in Step 2 into the trained deep multi-scale time-series coupled mapping network. Through the network, the operating condition-specific prior laws and multi-dimensional spatial coupling relationships are implicitly extracted in the mapping space. Time-series autoregressive evolution decoding is performed to output the day-ahead continuous hourly electricity price prediction curves and obtain the final day-ahead electricity price prediction value.

2. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: Collaboratively collect the historical day-ahead electricity price time-series data, and simultaneously collect multi-dimensional power operation related element data that are strongly correlated with the day-ahead electricity price; the multi-dimensional power operation related element data includes: day-ahead supply and demand time-series data, real-time key operation data, and day-ahead meteorological time-series data; Steps 1-2 involve performing a full-dimensional integrity retrieval on the collected multidimensional time-series data to accurately locate missing time-series positions and statistically analyze the span of time-series faults. A multidimensional time-series dynamic interpolation algorithm is then used to construct an interpolation benchmark based on historically known time-series feature points before and after the missing positions. Smooth interpolation is then performed on all missing values ​​in the multidimensional power operation related element data and electricity price time series to complete the data, outputting a standard electricity price time series and a multidimensional operating condition auxiliary feature dataset.

3. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 2, characterized in that, The day-ahead supply and demand time-series data includes: day-ahead direct dispatch load, day-ahead tie-line receiving load, day-ahead wind power, day-ahead photovoltaic power, day-ahead nuclear power, and day-ahead non-market-based generating unit power generation data; the real-time key operation data includes: real-time direct dispatch load, real-time tie-line receiving load, real-time wind power, real-time photovoltaic power, real-time nuclear power, and real-time non-market-based generating unit power generation data; the day-ahead meteorological time-series data includes: real-time temperature, real-time solar irradiance, and real-time wind speed.

4. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 1, characterized in that, In step 2, the nonlinear correlation contribution is calculated using an objective weighting algorithm, specifically as follows: The maximum mutual information coefficient is used to calculate the time series of the i-th auxiliary feature in the benchmark full sample dataset. Maximum mutual information coefficient between the current day electricity price series y and the current day electricity price series y : In the formula, For characteristic variables Mutual information with the electricity price variable y; and These represent the number of rows and columns into which the variable is divided into grids, respectively; B is the upper limit of the grid division. The obtained Normalization is performed to obtain the dynamic feature weights of the i-th auxiliary feature. : In the formula, m is the total number of auxiliary features.

5. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 4, characterized in that, In step 2, the grey relational analysis method is used to calculate the global macro-trend correlation degree. The calculation formula is as follows: In the formula, The overall similarity of the global macro trend between the date to be tested and the k-th historical candidate date. Let i be the dynamic feature weights of the auxiliary features. The grey relational coefficient between the k-th candidate day and the i-th auxiliary feature is calculated as follows: In the formula, It is the minimum difference between two levels. Let be the value of the i-th auxiliary feature for the day to be tested. Let be the value of the k-th historical candidate day on the i-th auxiliary feature; The maximum difference between the two levels; The resolution coefficient.

6. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 5, characterized in that, In step 2, the group confidence distance is thresholded using an asymmetric adaptive gating mechanism to determine the final similarity. This specifically includes the following steps: Step 2-1: Based on the available predictable sequences, extract statistical features to construct the working condition fingerprint vector for the day to be tested and the historical candidate day k working condition fingerprint vectors; after standardizing the working condition fingerprint vectors, calculate the spatial Euclidean distance between the standardized working condition fingerprint of the day to be tested and the working condition fingerprints of each historical candidate day. The operating condition similarity weights for the k-th historical candidate day are calculated using an exponential kernel function. ; Step 2-2: Calculate the spatial Euclidean distance between all historical candidate days and the day to be tested, and extract the set confidence quantile boundary values. ; Steps 2-3 introduce the global safety condition maximum threshold boundary. and adaptive gated switch variables Compare quantile boundary values Boundary to the maximum threshold of global safety conditions : If the relation is satisfied If the current working condition under test has sufficient clustering determinism in historical space, then the adaptive gating switch variable will be determined. Set to 1; If the relation is satisfied If the current test condition is determined to be a historically rare and unique scenario, the adaptive zero-axis collapse of the gating mechanism is triggered, and the adaptive gating switch variable is changed. Set to 0; Steps 2-4: Correlation of global macro trends Weight of similar working conditions and adaptive gating switch variables Perform global mathematical composition to construct the final similarity score. : 。 7. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 6, characterized in that, In step 2-1, the statistical features of the operating condition fingerprint vector include one or more combinations of the following: predicted load peak / valley / average, predicted load maximum ramp rate, predicted wind and solar power output average / fluctuation, wind and solar power penetration rate, and predicted temperature extreme values.

8. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 1, characterized in that, Step 3 specifically includes: Step 3-1: Establish a multi-scale parallel convolutional feature extraction module for the historical feature encoding branch. By setting multiple parallel convolutional paths with different kernel sizes and different dilation rates, the module extracts and concatenates the input historical information feature matrix to output a fused feature vector. Step 3-2: Establish a BiLSTM bidirectional temporal modeling module for the historical feature encoding branch, receive the fused feature vector, capture the temporal bidirectional dependency through forward and backward LSTM operation units, and output the splicing result of the bidirectional hidden layer state; Step 3-3: Establish the Attention filtering module for the historical feature encoding branch, and use the query, key, and value linear matrix to score and normalize the bidirectional temporal hidden layer state, and finally generate the historical feature context vector. Steps 3-4: Establish a BiLSTM bidirectional temporal modeling module for the predictive feature encoding branch, receive the predictive information feature matrix, extract the temporal evolution pattern of the daily predicted data sequence through forward and backward LSTM units, and output the predictive hidden layer state vector. Steps 3-5: Establish a feature splicing and fusion layer, splicing and combining the historical feature context vector and the predicted hidden layer state vector in the feature dimension to generate a joint high-dimensional feature vector; Steps 3-6: Establish the LSTM self-looping module of the decoder, perform autoregressive decoding on the joint high-dimensional features, and output the hidden layer state of the decoder. Steps 3-7 establish the reconstruction prediction module of the decoder. After performing random deactivation processing on the hidden layer state of the decoder, the predicted electricity price values ​​for consecutive hours before the day are output through the output linear layer mapping.

9. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 1, characterized in that, Step 3 also includes configuring the deep multi-scale temporally coupled mapping network: setting the continuous optimization hyperparameters of the deep multi-scale temporally coupled mapping network, using the mean absolute error as the loss function, and configuring the learning rate decay strategy using an adaptive moment estimation optimizer; wherein, the loss function and the learning rate decay mechanism are expressed as follows: In the formula, N is the total number of predicted samples; This represents the actual electricity price at time t. Here is the predicted electricity price at time t; MAE is the mean absolute error. Let be the learning rate for the t-th training round; The initial learning rate; This is the learning rate decay factor; This is the decay step size; This is the floor function.

10. The day-ahead electricity price forecasting method based on multi-dimensional operating condition adaptive gating according to claim 8, characterized in that, In step 4, the feature data of the day to be predicted is input into the trained deep multi-scale temporal coupled mapping network model for prediction, to obtain the final electricity price prediction result: In the formula, The historical information feature matrix of the day before the test date is used as the input to the historical feature encoding branch in the trained deep multi-scale temporal coupled mapping network model; The predicted information feature matrix for the day to be tested is used as the input for the predicted feature encoding branch; This represents a trained deep multi-scale temporally coupled mapping network model based on dual-branch feature fusion and encoder-decoder architecture. This is the predicted electricity price for the 24 hours prior to the test date, directly output by the model after dual-branch feature extraction, data splicing and fusion, and decoding mapping.

Images