Power market price forecasting method and device combining a timing large model and a fine-tuned model

Abstract

This application provides a method and apparatus for predicting electricity market prices by combining a large-scale time-series model and a fine-tuning model. The method includes collecting multi-source heterogeneous power time-series data from the power grid and performing normalization preprocessing to construct a unified data tensor containing batch dimension, time-series length dimension, and feature dimension. After performing hybrid positional encoding on the unified data tensor, multi-scale time-series correlation features are extracted through a multi-head attention mechanism, and time-series dependency modeling is performed in conjunction with a long short-term memory mechanism. High-dimensional cross features are constructed to obtain initial prediction results, and then the final electricity market price or power parameter prediction results are obtained by correcting them through a residual model. This application achieves efficient structured processing of multi-source heterogeneous data, extracts multi-scale time-series features of power system operation, effectively improves the accuracy and real-time performance of power parameter calculation, and can meet the requirements of real-time operation control of power systems.

Description

Technical Field

[0001] This application relates to the field of power technology, specifically to a method and apparatus for predicting electricity market prices by combining a large-scale time-series model and a fine-tuning model. Background Technology

[0002] With the continuous advancement of the construction of a unified power system, power system operation data exhibits significant characteristics of being ultra-large-scale, multi-source, heterogeneous, and strongly time-series correlated. Among these, power system operation parameters are jointly affected by fluctuations in new energy output, changes in grid topology, evolution of meteorological conditions, and multi-timescale operating mechanisms, exhibiting complex characteristics of high-frequency random fluctuations and multiple superimposed cycles. Their variation patterns are difficult to accurately describe using traditional physical models.

[0003] In existing methods for calculating power system operating parameters, traditional machine learning algorithms can handle some nonlinear relationships, but they have significant limitations when processing long-span time-series data, making it difficult to fully capture the complex dynamic evolution of system operation. While deep learning-based computational methods have improved data processing capabilities to some extent, they still suffer from insufficient extraction of multi-scale time-series features, poor fusion of different types of features, and insufficient stability of computational results, failing to meet the requirements of real-time operation control of power systems for parameter calculation accuracy and response speed.

[0004] In summary, existing technologies for calculating power system operating parameters suffer from technical problems such as insufficient processing capabilities for long-term, multi-source, heterogeneous data, inadequate multi-scale feature extraction, and difficulty in balancing computational accuracy and real-time performance. Summary of the Invention

[0005] This application addresses the problems existing in the prior art by providing a method and apparatus for predicting electricity market prices by combining a large-scale time-series model and a fine-tuning model, thereby solving the aforementioned problems.

[0006] To achieve the above objectives, the technical solution adopted in this application is as follows: On the one hand, this application provides a method for predicting electricity market prices by combining a large-scale time-series model and a fine-tuning model, which includes the following steps: Collect multi-source heterogeneous power time-series data from the power system; The multi-source heterogeneous power time-series data are subjected to normalization preprocessing to obtain standardized feature data; Based on the standardized feature data, a unified data tensor containing batch dimension, time series length dimension and feature dimension is constructed; The unified data tensor is subjected to hybrid positional encoding to obtain a temporal tensor; Multi-head attention mechanism is used to extract features from the temporal tensor to obtain multi-scale temporal correlation features; The temporal tensor is input into a preset temporal large model; the preset temporal large model adopts a hybrid architecture of Transformer and convolutional neural network. In the preset time-series large model, the multi-scale time-series correlation features are modeled according to the long short-term memory mechanism to obtain time-dependent features; the time-dependent features are cross-combined and nonlinearly transformed to construct high-dimensional cross features; based on the high-dimensional cross features, the initial prediction results are obtained; the initial prediction results are corrected by the residual model to obtain the electricity market price prediction results.

[0007] Optionally, the multi-source heterogeneous power time-series data includes power system operation data, meteorological data, and time characteristic data; The time feature data includes: Obtain the raw date and time data; The original date and time data are transformed by performing trigonometric functions on the four independent time dimensions of month, date, weekday, and hour to generate the time feature data.

[0008] Optionally, the normalization preprocessing includes: Extract each type of feature data from the multi-source heterogeneous power time-series data; Perform an independent normalization operation on each type of feature data; In the independent normalization operation, each type of feature data is mapped to the [0,1] value range to obtain the standardized feature data.

[0009] Optionally, the hybrid position encoding is a rotational position encoding that fuses a sine function; The hybrid position encoding includes: Extract the data for each time point from the unified data tensor; For each time point data, both absolute position signals and relative position signals are simultaneously embedded; The absolute position signal is generated by a sine function, and the relative position signal is generated by rotational position encoding. The temporal tensor containing the embedded location information is obtained through hybrid encoding.

[0010] Optionally, the multi-head attention mechanism includes at least three sets of parallel attention heads, including a first set of attention heads, a second set of attention heads, and a third set of attention heads; The process of obtaining the multi-scale temporal correlation features includes: Using the first set of attention heads, short-term fluctuation features are extracted from the time-series tensor of the embedded location information, and the short-term fluctuation features correspond to the power parameter change patterns within 1 to 3 days. Using the second set of attention heads, periodic patterns are extracted from the time-series tensor containing the embedded location information. These periodic patterns correspond to the periodic changes in daily, weekly, monthly, and annual power parameters. Using the third set of attention heads, abnormal correlation features are extracted from the time-series tensor of the embedded location information. These abnormal correlation features correspond to parameter changes under sudden power system conditions. The abnormal correlation features include at least holiday electricity price jump features, electricity price abrupt change features under supply and demand imbalance, and electricity price abnormal fluctuation features caused by grid congestion. By integrating the short-term fluctuation characteristics, periodic pattern characteristics, and abnormal correlation characteristics, the multi-scale time series correlation characteristics are obtained.

[0011] Optionally, the construction of the preset time-series large model includes: Obtain the actual values ​​of historical power parameters as tag data; The label data and feature variables are simultaneously input into the time series large model for joint training, and the preset time series large model is obtained after training.

[0012] Optionally, the training process of the preset time-series large model applies a combination of early stopping mechanism and learning rate decay mechanism; The early stopping mechanism is used to monitor changes in the validation set loss function; if the validation set loss does not decrease within a preset number of consecutive rounds, the training process of the preset time series large model is terminated. In the learning rate decay mechanism, cosine annealing is used to adjust the learning rate, which includes reducing the learning rate as the number of training rounds increases.

[0013] Optionally, correcting the initial prediction result includes: Based on the historical output residual dataset of the preset time series large model, the residual model is pre-trained and obtained; The residual model fits the residual between the output value and the true value of the preset time series large model to obtain the residual correction value; The initial prediction result is superimposed with the residual correction value to obtain the electricity market price prediction result.

[0014] On the other hand, this application provides a method for predicting power parameters, including: Collect multi-source heterogeneous power time-series data from the power system; The multi-source heterogeneous power time-series data are subjected to normalization preprocessing to obtain standardized feature data; Based on the standardized feature data, a unified data tensor containing batch dimension, time series length dimension and feature dimension is constructed; The unified data tensor is subjected to hybrid positional encoding to obtain a temporal tensor; Multi-head attention mechanism is used to extract features from the temporal tensor to obtain multi-scale temporal correlation features; Based on the long short-term memory mechanism, the multi-scale temporal correlation features are modeled for temporal dependence to obtain temporal dependence features; The time-dependent features are cross-combined and nonlinearly transformed to construct high-dimensional cross features; Linear projection is performed on the high-dimensional cross features to obtain the initial prediction result; Based on the high-dimensional cross features, the initial prediction results are corrected using a residual model to obtain the power parameter prediction results.

[0015] Furthermore, this application also provides a power parameter prediction device, which mainly includes a data processing module, an encoding module, a feature extraction module, a high-dimensional feature construction module, a main prediction module, and a fine-tuning module; The data processing module receives multi-source heterogeneous power time-series data, processes the multi-source heterogeneous power time-series data, and outputs a unified data tensor. The encoding module receives the unified data tensor, encodes the unified data tensor, and outputs a temporal tensor with embedded position information. The feature extraction module receives the temporal tensor containing the embedded location information, and performs feature extraction on the temporal tensor containing the embedded location information and outputs temporal dependent features. The high-dimensional feature construction module receives the temporal dependency features, and is used to construct features from the temporal dependency features and output high-dimensional cross features. The main prediction module receives the high-dimensional cross features, calculates the high-dimensional cross features, and outputs the initial prediction results. The fine-tuning module is configured with a residual training model based on the Transformer architecture to train the relationship between the residuals of the large model prediction results and the actual results and the features constructed by the large model. For new prediction data, a time series large model is used to predict the results, and the trained residual model is called to perform residual prediction in advance. The initial prediction results are superimposed with the residual correction values ​​to output the electricity market price prediction results or the electricity parameter prediction results. The data processing module performs normalization preprocessing on the multi-source heterogeneous power time-series data to obtain standardized feature data; The data processing module also constructs the unified data tensor based on the standardized feature data.

[0016] Compared with the prior art, this application has the following advantages: This application achieves efficient structured processing of multi-source heterogeneous power time-series data by constructing a unified data tensor. Then, by combining hybrid positional encoding with a multi-head attention mechanism, it fully extracts the multi-scale time-series correlation features of power system operation. By combining a hybrid architecture of Transformer and convolutional neural network with a large time-series model and residual correction, it effectively improves the accuracy and real-time performance of electricity market price prediction and power parameter calculation, which can meet the requirements of real-time operation control of power system. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the method in this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] It should be noted that, for ease of understanding, the method steps in the specific embodiments of this application are described in a certain order, but those skilled in the art can change the order of the steps according to actual needs, so this should not be used as a limiting condition; further, in the description of the following specific embodiments, the superscripts and subscripts of each parameter should be understood as distinguishing marks of similar identifiers in accordance with common interpretations unless otherwise specified, representing the parameters of the related or corresponding devices, and should not be understood as specific models or special marks.

[0022] like Figure 1As shown, this application provides a method for predicting electricity market prices by combining a large-scale time-series model and a fine-tuning model. The electricity market price referred to in this application includes, but is not limited to, nodal marginal prices. This method can also predict electricity parameters, which are parameters that characterize the operating state of the power system, including but not limited to system operating parameters and electricity cost parameters. Taking nodal marginal prices as an example, these are core physical parameters reflecting the power balance and transmission congestion state of the power grid. Their values ​​are determined by the physical operating laws of the power grid. The method specifically includes the following steps: First, collect multi-source heterogeneous power time-series data from the power system.

[0023] Furthermore, multi-source heterogeneous power time-series data includes power system operation data, meteorological data, and time characteristic data. Power system operation data corresponds to publicly disclosed market data, specifically including system load, tie-line exchange power, non-market output, wind power output, photovoltaic output, total renewable energy output, hydropower output, mandatory start / stop unit capacity, total generation output, and system adjustable output space. System adjustable output space refers to the range of output adjustment that thermal power units need to provide to maintain grid power balance; its equivalent representation can be converted into thermal power bidding space. The aforementioned power system operation data corresponds to publicly disclosed physical operation data in the power market.

[0024] This application introduces extended features such as system load factor, the difference between new energy sources and system load, and the system power supply-demand ratio. The system power supply-demand ratio refers to the ratio of the total power generation capacity of the power grid to the total power load, and its physical meaning is consistent with the market supply-demand ratio, reflecting the power balance state of the power grid. Therefore, the above-mentioned introduction can better reflect the power system supply and demand situation and reduce the impact of anomalies in a single data source on the calculation results. Meteorological data covers more than 30 sub-indicators, including wind direction and speed at 100m and 10m altitudes, 2m dew point temperature, 2m temperature, skin temperature, net solar radiation at the surface, downward solar radiation at the surface, and total precipitation.

[0025] For time-related data, the original date and time data is obtained, and then trigonometric function transformations are performed on the original date and time data for the four independent time dimensions of month, day, week, and hour to generate time-related data. Specifically, one example is to transform the month data separately. Monthly specific date data Weekly data and point-in-time data Perform the following processing: ; ; ; ; In the formula, The sine characteristic of the month, The cosine characteristic of the month; For the specific dates of each month, the sine characteristic is... Cosine characteristics for specific dates each month; Characterized by the sine wave pattern of the week. This is a characteristic of the week's cosine. The sinusoidal characteristic at time points, The data is represented by the cosine feature at a given time point; (1-12) represents the month number; (1-31) represents the specific date of each month; (0-6) represents the week number, with 0 representing Monday; (0-95) represents the specific time of each day, with 0 representing 00:00, 1 representing 00:15, and so on, with 95 representing 23:45, corresponding to a 15-minute data sampling interval.

[0026] Finally, a time feature set containing hour, year, month, day and corresponding trigonometric function continuous transformation features is constructed, namely the aforementioned time feature data.

[0027] Multi-source heterogeneous power time-series data are subjected to normalization preprocessing to obtain standardized feature data. The normalization preprocessing includes: Each type of feature data is extracted from multi-source heterogeneous power time-series data, and an independent normalization operation is performed on each type of feature data. In the independent normalization operation, each type of feature data is mapped to the [0,1] numerical range to obtain standardized feature data. Specifically, this application takes the min-max normalization method as an example for normalization preprocessing, and the normalization formula is as follows: ; In the formula, These are the normalized feature data values ​​after sampling period d and sampling time t; These are the characteristic data values ​​for the sampling period d and the sampling time t; The minimum value of the same type of feature data in the historical dataset; The maximum value of the same type of feature data in the historical dataset; d is the sampling period, and D is the total number of the same type of feature data in the historical dataset.

[0028] This method can avoid the bias caused by differences in the numerical range of feature data affecting the model's generalization ability, thereby accelerating model training and improving computational accuracy.

[0029] Based on standardized feature data, a unified data tensor is constructed, comprising batch dimension, time series length dimension, and feature dimension. In this application, basic data from 2 to 3 years of historical data every 15 minutes is selected, containing more than 300 features and a total of more than 32 million data points. This multivariate and heterogeneous time series data is then structured to form a unified data foundation, namely, a unified data tensor.

[0030] Hybrid positional encoding is applied to a unified data tensor to obtain a temporal tensor. Hybrid positional encoding is a rotational positional encoding that fuses a sine function. Specifically, hybrid positional encoding includes: Extract the data at each time point from the unified data tensor, and embed both absolute and relative position signals into the data at each time point. The absolute position signal is generated using a sine function, and the relative position signal is generated using rotational position encoding. A temporal tensor embedding position information is obtained through hybrid encoding. Specifically, the hybrid position encoding used in this application involves adding the absolute position signal generated by the sine position encoding and the relative position signal generated by the rotational position encoding element-wise to obtain the final position encoding vector.

[0031] The encoding method proposed in this application can better capture the contextual relationships of tensor data, laying the foundation for multi-head parallel data processing and model computation.

[0032] A pre-built and trained temporal large-scale model is constructed and trained based on a hybrid architecture of Transformer and Convolutional Neural Network (CNN). Specifically, the pre-built temporal large-scale model adopts a serially connected hybrid architecture. Temporally dependent features are first input into the CNN to extract local spatial features, and then the output of the CNN is input into the Transformer encoder to extract global temporal features.

[0033] The convolutional neural network is further optimized based on the Convolutional Positional Encoding (Conv Positional Encoding) and Convolutional Feedforward Network (ConvFFN) structures. It adopts a pure channel-wise separable Conv1d structure, removing the pointwise convolutions from the original architecture for extreme lightweighting. All convolutional layers have a uniform kernel size of 3, a stride of 1, and the same padding method, corresponding to a receptive field of one time step before and after (a local window of three time steps). The Conv Positional Encoding layer has 12 groups, and the Convolutional Feedforward network layer has 512 groups. No additional batch normalization layer follows the convolutional layers; they are directly connected to subsequent linear layers.

[0034] Before the features are input into the Transformer encoder, they are first mapped to 128 dimensions through an input projection layer. The Transformer employs a multi-branch attention architecture, with each attention branch containing an encoder block consisting of two layers, each with four attention heads. The model's hidden layers have a dimension of 128, and the feedforward neural network has a dimension of 512. The activation function is uniformly GELU, and the normalization method is uniformly Layer Norm. The encoder output is finally mapped to 96 dimensions through a predictor layer, corresponding to the prediction results for the next 96 time steps.

[0035] Furthermore, the preset time-series large-scale model has an optimizable parameter scale of 10 million to 15 million. The model simultaneously utilizes labels and feature variables, along with representations learned from historical data, for joint computation. Specifically, one example uses a preset time-series large-scale model based on the TimesFM-2.5 architecture, and employs the following core parameter configurations to address the strong nonlinearity and multi-scale fluctuation characteristics of power system time-series data: It adopts a Decoder-only pure decoder architecture, which contains 20 Transformer hidden layers, each layer has 16 attention heads, each head has a dimension of 80, the total hidden layer dimension is 1280, the feedforward neural network dimension is 1280, and the activation function is the swish(SiLU) function; The maximum sequence length is 16384, the context length is 16384, the prediction length is 128, and the block length is 32; the normalization method used is RMS Norm, ε=10. -6 The position encoding uses RoPE rotational position encoding, θ=10000.

[0036] It should be noted that the above parameters are an optimized configuration for the electricity market price forecasting task. Since there are differences in data distribution, time series length and forecasting targets in different application scenarios, those skilled in the art can make adaptive adjustments to the above parameters according to the data characteristics of the actual project, and all can achieve the technical effects of this application.

[0037] The training of the pre-defined time-series large model includes: The true values ​​of historical power parameters are obtained as label data. The label data and feature variables are simultaneously input into the time series large model for joint training, and the preset time series large model is obtained after training.

[0038] The training process employs a combination of early stopping and learning rate decay. Early stopping monitors changes in the validation set loss function; if the validation set loss does not decrease within a preset number of training epochs, the training of the large time series model is terminated, effectively preventing overfitting. The learning rate decay mechanism uses cosine annealing to adjust the learning rate. One example uses the AdamW optimizer with an initial learning rate of 1e-4, a weight decay coefficient of 1e-5, and a batch size of 64. The number of consecutive training epochs without a decrease in the early stopping rate is set to 10, and the cosine annealing decay period is set to 50 training epochs. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio. These training parameters represent an optimal configuration for the power time series prediction task. Those skilled in the art can also make adaptive adjustments based on model size and dataset size. Specific adjustments include reducing the learning rate as the training epochs increase to help the model more precisely approximate the optimal solution while maintaining a relatively fast convergence speed in the early stages of training.

[0039] Therefore, in the pre-defined temporal model of this application, a multi-head attention mechanism is used to extract features from the aforementioned temporal tensor to obtain multi-scale temporal correlation features. The multi-head attention mechanism of this application includes at least three sets of parallel attention heads; one example includes a first set of attention heads, a second set of attention heads, and a third set of attention heads. Thus, the process of obtaining multi-scale temporal correlation features includes: Using the first set of attention heads, short-term fluctuation features are extracted from the time-series tensor with embedded location information. These short-term fluctuation features correspond to the power parameter variation patterns within 1 to 3 days. Using the second set of attention heads, periodic features are extracted from the time-series tensors containing embedded location information. These periodic features correspond to the periodic changes in power parameters over days, weeks, months, and years. Using a third set of attention heads, abnormal correlation features are extracted from the time-series tensors embedded with location information. These abnormal correlation features correspond to parameter changes under sudden power system conditions. Specifically, the abnormal correlation features include at least the characteristics of electricity price jumps during holidays, the characteristics of sudden electricity price changes under supply-demand imbalances, and the characteristics of abnormal electricity price fluctuations caused by grid congestion.

[0040] It's understandable that the price jumps during holidays reflect the price fluctuations during holidays. For example, from Monday to Friday, the electricity market price is between 0.3 and 0.4 yuan / kWh. During holidays, as electricity demand decreases, the price may drop to 0.1-0.2 yuan / kWh. Therefore, the aforementioned focus is used to identify price fluctuation patterns between non-holiday and holiday periods. The sudden price fluctuations under supply-demand imbalances reflect the price changes under conditions of supply-demand imbalance. For example, during certain periods, there is a large amount of renewable energy generation (such as photovoltaic power) but low electricity demand, resulting in oversupply and lower prices. Conversely, during other periods, renewable energy output is low while load demand is high, resulting in undersupply and higher prices. Abnormal price fluctuations caused by grid congestion refer to congestion on a transmission line, leading to higher prices at the load nodes supplied by that line. Apart from these special cases, electricity prices are generally within the normal range at other times.

[0041] The three attention heads use the same network structure and extract features at different scales by setting different attention window sizes. The first attention head uses a window size of n time steps, the second attention head uses a window size of m time steps, and the third attention head uses a windowless global attention mechanism. The multi-scale temporal correlation features are obtained by fusing short-term fluctuation features, periodic regularity features, and abnormal correlation features. Here, n and m are both positive integers.

[0042] Multiple attention heads work in parallel to capture different patterns. The differentiated information extracted by each head is fused to form a comprehensive contextual representation, which inherently reveals the influence weights and correlations among various features.

[0043] It is important to distinguish that the model in this application has three sets of parallel independent attention branches, which correspond to three feature extraction tasks: short-term fluctuations, periodic patterns, and abnormal correlations. Each branch uses an independent two-layer Transformer encoder, with each layer containing four standard self-attention heads. This architecture can extract multi-scale temporal features in parallel while ensuring the modeling accuracy of each scale feature.

[0044] Based on the Long Short-Term Memory (LSTM) mechanism, time-series dependency models are performed on multi-scale temporal correlation features to obtain time-series dependency features. In the LSTM mechanism, short-term memory focuses on the high-frequency fluctuation patterns of power parameters over multiple days, while long-term memory grasps the macroscopic changes in power parameters, such as trends and seasonality.

[0045] This invention constructs high-dimensional cross-features by cross-combining and nonlinearly transforming temporally dependent features. This hybrid architecture enables the cross-combination and nonlinear transformation of features, constructing more predictive high-order features. The serially connected hybrid architecture employed in this application first inputs the temporally dependent features into a convolutional neural network to extract local spatial features, and then inputs the output of the convolutional neural network into a Transformer encoder to extract global temporal features.

[0046] The initial prediction result is obtained by using high-dimensional cross features, that is, by using a linear prediction layer (Predictor) to map the high-dimensional cross features to the target prediction length dimension, thereby obtaining the initial prediction result.

[0047] The initial prediction results are corrected using a residual model to obtain the electricity market price prediction results. Correcting the initial prediction results involves constructing a residual model of the large-scale model calculation results and the actual values, and then training and predicting them separately. One example in this application is that the residual model uses a neural network structure and the ReLU function as the activation function. Furthermore, during the training phase of the residual model, reasonable forward propagation, loss, and gradient update functions are designed to determine the scale of the large-scale model calculation values ​​and the actual values, and the training model is then saved. Faced with new data, during the prediction phase of the residual model, a pre-trained residual model is invoked, and the aforementioned high-dimensional cross features are input to obtain the prediction residuals. These prediction residuals are then used to fine-tune the new computational results of the large-scale time-series model. The aforementioned residual model correction process improves accuracy by fitting the computational error of the physical parameters; specifically, based on the historical output residual dataset of the pre-defined large-scale time-series model, the residual model is pre-trained and obtained.

[0048] The residual model fits the residual between the output value and the actual value of the preset time series large model to obtain the residual correction value. The initial prediction result and the residual correction value are superimposed to obtain the electricity market price prediction result.

[0049] Due to the strong nonlinearity, non-stationarity, and sudden fluctuations in electricity prices in the power market, a single large-scale time series model cannot simultaneously capture both global trend prediction and local anomaly detection. This application constructs an independent residual adjustment model to fit the local price fluctuation characteristics and prediction error patterns that the large-scale time series model cannot cover, significantly improving prediction accuracy under extreme conditions. The residual model and the main large-scale time series model adopt a parallel training and joint prediction working mode. The main model is responsible for learning the long-term trend and cyclical patterns of electricity prices, while the residual model is responsible for correcting the prediction bias of the main model. The two work together to achieve high-precision electricity price prediction.

[0050] On the other hand, this application also provides a power parameter prediction device, including a data processing module, an encoding module, a feature extraction module, a high-dimensional feature construction module, a main prediction module, and a fine-tuning module.

[0051] The modules are as follows: The data processing module receives multi-source heterogeneous power time-series data, processes it, and outputs a unified data tensor. The module performs normalization preprocessing on the multi-source heterogeneous power time-series data to obtain standardized feature data, and then constructs a unified data tensor based on this standardized feature data.

[0052] The encoding module receives a uniform data tensor, encodes the uniform data tensor, and outputs a temporal tensor with embedded position information.

[0053] The feature extraction module receives a temporal tensor with embedded location information, and uses it to extract features from the temporal tensor with embedded location information and output temporally dependent features.

[0054] The high-dimensional feature construction module receives temporally dependent features, uses them to construct features from the temporally dependent features, and outputs high-dimensional cross features.

[0055] The main prediction module receives high-dimensional cross features, calculates the high-dimensional cross features, and outputs the initial prediction results.

[0056] The fine-tuning module is configured with a residual training model based on the Transformer architecture to train the relationship between the residuals of the large model prediction results and the actual results and the features constructed by the large model. For new prediction data, the time series large model is used to predict the results, and the trained residual model is called to perform residual prediction in advance. The initial prediction results are superimposed with the residual correction values ​​to output the electricity market price prediction results or the electricity parameter prediction results.

[0057] Finally, it should be noted that the above content is only used to illustrate the technical solution of this application, and is not intended to limit the scope of protection of this application. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of this application shall not depart from the substance and scope of the technical solution of this application.

Claims

1. A method for predicting electricity market prices by combining a large-scale time-series model and a fine-tuning model, characterized in that, include: Collect multi-source heterogeneous power time-series data from the power system; The multi-source heterogeneous power time-series data are subjected to normalization preprocessing to obtain standardized feature data; Based on the standardized feature data, a unified data tensor containing batch dimension, time series length dimension and feature dimension is constructed; The unified data tensor is subjected to hybrid positional encoding to obtain a temporal tensor; The temporal tensor is input into a preset temporal large model; the preset temporal large model adopts a hybrid architecture of Transformer and convolutional neural network. In the preset temporal large model, the temporal tensor is used to extract features through a multi-head attention mechanism to obtain multi-scale temporal correlation features; Based on the long short-term memory mechanism, the multi-scale temporal correlation features are modeled for temporal dependence to obtain temporal dependence features; The time-dependent features are cross-combined and nonlinearly transformed to construct high-dimensional cross features; Based on the high-dimensional cross features, the initial prediction result is obtained; The initial prediction results are corrected using a residual model to obtain the electricity market price prediction results.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous power time-series data includes power system operation data, meteorological data, and time characteristic data; The time feature data includes: Obtain the raw date and time data; The original date and time data are transformed by performing trigonometric functions on the four independent time dimensions of month, date, weekday, and hour to generate the time feature data.

3. The method according to claim 2, characterized in that, The normalization preprocessing includes: Extract each type of feature data from the multi-source heterogeneous power time-series data; Perform an independent normalization operation on each type of feature data; In the independent normalization operation, each type of feature data is mapped to the [0,1] value range to obtain the standardized feature data.

4. The method according to claim 1, characterized in that, The hybrid position encoding is a rotational position encoding that fuses a sine function; The hybrid position encoding includes: Extract the data for each time point from the unified data tensor; For each time point data, both absolute position signals and relative position signals are simultaneously embedded; The absolute position signal is generated by a sine function, and the relative position signal is generated by rotational position encoding. The temporal tensor containing the embedded location information is obtained through hybrid encoding.

5. The method according to claim 4, characterized in that, The multi-head attention mechanism includes at least three sets of parallel attention heads, including a first set of attention heads, a second set of attention heads, and a third set of attention heads; The process of obtaining the multi-scale temporal correlation features includes: Using the first set of attention heads, short-term fluctuation features are extracted from the time-series tensor of the embedded location information, and the short-term fluctuation features correspond to the power parameter change patterns within 1 to 3 days. Using the second set of attention heads, periodic patterns are extracted from the time-series tensor containing the embedded location information. These periodic patterns correspond to the periodic changes in daily, weekly, monthly, and annual power parameters. Using the third set of attention heads, abnormal correlation features are extracted from the time-series tensor of the embedded location information. These abnormal correlation features correspond to parameter changes under sudden power system conditions. The abnormal correlation features include at least holiday electricity price jump features, electricity price abrupt change features under supply and demand imbalance, and electricity price abnormal fluctuation features caused by grid congestion. By integrating the short-term fluctuation characteristics, periodic pattern characteristics, and abnormal correlation characteristics, the multi-scale time series correlation characteristics are obtained.

6. The method according to claim 1, characterized in that, The construction of the preset time series large model includes: Obtain the actual values ​​of historical power parameters as tag data; The label data and feature variables are simultaneously input into the time series large model for joint training, and the preset time series large model is obtained after training.

7. The method according to claim 6, characterized in that, The training process of the pre-defined time-series large model applies a combination of early stopping mechanism and learning rate decay mechanism; The early stopping mechanism is used to monitor changes in the loss function on the validation set; If the validation set loss does not decrease within a consecutive preset number of rounds, the training process of the preset time series large model is terminated. In the learning rate decay mechanism, cosine annealing is used to adjust the learning rate, which includes reducing the learning rate as the number of training rounds increases.

8. The method according to claim 1, characterized in that, The correction of the initial prediction results includes: Based on the historical output residual dataset of the preset time series large model, the residual model is pre-trained and obtained; The residual model fits the residual between the output value and the true value of the preset time series large model to obtain the residual correction value; The initial prediction result is superimposed with the residual correction value to obtain the electricity market price prediction result.

9. A method for predicting power parameters, characterized in that, include: Collect multi-source heterogeneous power time-series data from the power system; The multi-source heterogeneous power time-series data are subjected to normalization preprocessing to obtain standardized feature data; Based on the standardized feature data, a unified data tensor containing batch dimension, time series length dimension and feature dimension is constructed; The unified data tensor is subjected to hybrid positional encoding to obtain a temporal tensor; Multi-head attention mechanism is used to extract features from the temporal tensor to obtain multi-scale temporal correlation features; Based on the long short-term memory mechanism, the multi-scale temporal correlation features are modeled for temporal dependence to obtain temporal dependence features; The time-dependent features are cross-combined and nonlinearly transformed to construct high-dimensional cross features; Linear projection is performed on the high-dimensional cross features to obtain the initial prediction result; Based on the high-dimensional cross features, the initial prediction results are corrected using a residual model to obtain the power parameter prediction results.

10. A power parameter prediction device, characterized in that, It includes a data processing module, an encoding module, a feature extraction module, a high-dimensional feature construction module, a main prediction module, and a fine-tuning module; The data processing module receives multi-source heterogeneous power time-series data, processes the multi-source heterogeneous power time-series data, and outputs a unified data tensor. The encoding module receives the unified data tensor, encodes the unified data tensor, and outputs a temporal tensor with embedded position information. The feature extraction module receives the temporal tensor containing the embedded location information, and performs feature extraction on the temporal tensor containing the embedded location information and outputs temporal dependent features. The high-dimensional feature construction module receives the temporal dependency features, and is used to construct features from the temporal dependency features and output high-dimensional cross features. The main prediction module receives the high-dimensional cross features, calculates the high-dimensional cross features, and outputs the initial prediction results. The fine-tuning module is configured with a residual training model based on the Transformer architecture. The residual training model is pre-trained to train the relationship between the residuals of the large model prediction results and the actual results and the features constructed by the large model. For new prediction data, the time series large model is used to predict the results, the trained residual model is called to perform residual prediction, the initial prediction results are superimposed with the residual correction values, and the electricity market price prediction results or electricity parameter prediction results are output. The data processing module performs normalization preprocessing on the multi-source heterogeneous power time-series data to obtain standardized feature data; The data processing module also constructs the unified data tensor based on the standardized feature data.

Images