Coal-fired unit short-term load prediction method, system, device and storage medium

Abstract

The application discloses a coal-fired unit short-term load prediction method, system, device and storage medium, belongs to the technical field of coal-fired unit short-term load prediction, obtains historical operation data of a coal-fired unit and meteorological characteristic data influencing load, constructs a DTA-Autoformer prediction model, embeds a dynamic time sequence attention module between an input projection layer and an encoder, captures sequence context dependence by using a bidirectional GRU, adaptively generates time step weights, and combines residual connection to enhance local feature expression; the prediction model is iteratively trained by using a training set and a validation set, the prediction performance of the prediction model is evaluated by a test set, unit operation and meteorological characteristic data are input into the trained prediction model, and a coal-fired unit short-term load prediction result is generated; the application effectively fuses the autocorrelation of Autoformer and the local dynamic feature extraction capability of the dynamic time sequence attention module, and improves the precision and robustness of coal-fired unit short-term load prediction.

Description

Technical Field

[0001] This invention belongs to the field of short-term load forecasting technology for coal-fired power units, specifically relating to methods, systems, equipment, and storage media for short-term load forecasting of coal-fired power units. Background Technology

[0002] Coal-fired power units are the core of modern power systems, bearing the heavy responsibility of ensuring the safe and stable operation of the power grid, flexible regulation, and providing a safety net for power supply. Under the background of building a new power system, the large-scale grid connection of new energy sources has exacerbated the challenges of source-load balance. Coal-fired power units need to frequently and deeply regulate peak loads and rapidly change loads, resulting in complex operating conditions. Their load sequences exhibit strong non-stationarity, nonlinearity, and time-varying characteristics. Furthermore, influenced by factors such as the physical characteristics of the units and fluctuations in coal quality, the operating data contains a large amount of noise and outliers, increasing the difficulty of high-precision load forecasting.

[0003] Existing short-term load forecasting methods are divided into two categories: traditional statistics and artificial intelligence. Both have limitations in dealing with complex scenarios: Traditional statistical methods (such as autoregressive integral moving average models and Kalman filtering) rely on the assumption of data stationarity or linearity, making it difficult to fit the nonlinear dynamic characteristics of units under varying operating conditions, resulting in insufficient accuracy; Recurrent Neural Networks (RNNs) and their variants are good at capturing time-series dependencies, but their serial computation efficiency is low, they are prone to gradient anomalies when processing long sequences, and they are sensitive to data quality and lack robustness against anomalies.

[0004] In recent years, prediction models based on the Transformer architecture have emerged. While they have solved the long-distance dependency problem and achieved parallel computing through self-attention mechanisms, the dot product attention mechanism of the standard Transformer ignores the inherent continuity and periodicity of time series during computation, often leading to inaccurate capture of trend information. Although the Autoformer model innovatively introduces an autocorrelation mechanism to extract the periodicity of the sequence, significantly improving the prediction effect of long series, it focuses too much on global periodicity. When faced with sudden, non-periodic, and drastic local fluctuations in coal-fired power unit loads caused by dispatching instructions (such as rapid ramp-up or load reduction phases), it often reacts slowly, exhibiting significant prediction lag or amplitude deviation. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of significant prediction lag and amplitude deviation in existing short-term load prediction methods for coal-fired power units, and thus proposes a short-term load prediction method, system, equipment and storage medium for coal-fired power units.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for short-term load forecasting of coal-fired power units, comprising the following steps: Historical operating data of coal-fired power units and meteorological characteristic data affecting load changes are obtained. The historical operating data and meteorological characteristic data are preprocessed to generate model input characteristic data, and the model input characteristic data are divided into training set, validation set and test set. A DTA-Autoformer prediction model is constructed by inputting the model's input feature data into the model and projecting the data onto a unified model dimension to generate a high-dimensional feature sequence. A bidirectional time-series representation is generated based on this high-dimensional feature sequence. Dynamic attention weights for each time step are generated based on the bidirectional time-series representation. These dynamic attention weights are weighted with the high-dimensional feature sequence and then added to the high-dimensional feature sequence through a residual connection domain to obtain a fused feature sequence. Autocorrelation encoding is performed on the fused feature sequence to generate encoded features. After cross-correlation interaction using the encoded features, trend and periodic terms are extracted and fused through sequence decomposition to generate a fused feature of trend and periodic terms. This fused feature of trend and periodic terms is mapped to the target dimension to generate a normalized load prediction sequence. The DTA-Autoformer prediction model is iteratively trained using the training and validation sets, and its prediction performance is evaluated using the test set to obtain the trained DTA-Autoformer prediction model. The historical operating data of coal-fired power units and meteorological characteristic data affecting load changes are input into the trained DTA-Autoformer prediction model. The output normalized load prediction sequence is then denormalized to generate short-term load prediction results for coal-fired power units.

[0007] Furthermore, the historical operational data and meteorological characteristic data are preprocessed, including: Outlier detection and cleaning, missing value imputation, feature selection based on Pearson correlation coefficient, time feature encoding, and normalization are performed on historical operational data and meteorological characteristic data to generate model input feature data.

[0008] Furthermore, the DTA-Autoformer prediction model sequentially includes an input projection layer, a dynamic temporal attention module, an Autoformer encoder, an Autoformer decoder, and an output prediction layer; The input projection layer projects the model input feature data onto a unified model dimension, generating a high-dimensional feature sequence. The dynamic temporal attention module includes a bidirectional gated recurrent unit, an attention scoring unit, and a residual fusion unit. The bidirectional gated recurrent unit receives the high-dimensional feature sequence. The attention scoring unit receives the output of the bidirectional gated recurrent unit and generates dynamic attention weights for each time step. The residual fusion unit multiplies the dynamic attention weights element-wise with the high-dimensional feature sequence and adds them to the high-dimensional feature sequence through a residual concatenation, outputting a fused feature sequence to the Autoformer encoder. The Autoformer encoder encodes the fused feature sequence. The Autoformer decoder receives the output of the Autoformer encoder and decodes it. The output prediction layer maps the output of the Autoformer decoder to the target dimension, generating a normalized load prediction sequence. The working process of the dynamic temporal attention module includes: The bidirectional gated cyclic unit receives high-dimensional feature sequences and outputs bidirectional time-series representations. The attention scoring unit receives bidirectional temporal representations through a two-layer fully connected network, generates raw attention scores for each time step, performs Softmax normalization on the raw attention scores in the time dimension, and generates dynamic attention weights for each time step. The residual fusion unit multiplies the dynamic attention weights element-wise with the high-dimensional feature sequence to obtain weighted features, and adds the weighted features with the high-dimensional feature sequence through residual connections to output the fused feature sequence.

[0009] Furthermore, the model input feature data includes: Features whose absolute value of the load correlation coefficient is greater than a preset limit are selected from historical operating data; Temperature, humidity, and rainfall data obtained from meteorological characteristic data; And hourly and weekday features extracted from the original timestamps and periodically encoded using sine and cosine.

[0010] Furthermore, the Autoformer encoder consists of multiple stacked encoder layers, each of which includes an autocorrelation mechanism module and a sequence decomposition module in sequence; The sequence decomposition module uses a moving average filter to decompose the input sequence into a trend term and a period term; The autocorrelation mechanism module calculates the autocorrelation function of the sequence through fast Fourier transform, selects the top-c significant delays, performs rolling alignment of the value sequence according to the selected delays, and aggregates them based on the autocorrelation values.

[0011] Furthermore, the Autoformer decoder consists of multiple stacked decoder layers, each of which sequentially includes a first autocorrelation mechanism module, an encoder-decoder interactive autocorrelation mechanism module, and a sequence decomposition module; The Autoformer decoder uses the latter half of the input sequence as a starting marker, concatenates zero vector placeholders to construct the decoder input, and initializes the decoder input as periodic and trend term inputs through the sequence decomposition module. In each decoder layer, the periodic term input is processed sequentially by the first autocorrelation mechanism module, the encoder-decoder interactive autocorrelation mechanism module, and the sequence decomposition module. The separated trend component is accumulated with the initial trend term input, and finally the periodic term prediction part is added with the accumulated trend prediction part to obtain the decoding result.

[0012] Furthermore, the DTA-Autoformer prediction model is iteratively trained using the training and validation sets, including: The mean squared error is used as the loss function to calculate the error between the predicted value and the true value. The AdamW optimizer is used to update the model parameters based on error backpropagation; An early stopping strategy is implemented based on the prediction error of the validation set. Training is stopped when the validation loss fails to improve for a preset number of consecutive rounds, and the optimal model parameters are saved.

[0013] Secondly, the present invention provides a short-term load forecasting system for coal-fired power units, comprising: The data processing module is used to acquire historical operating data of coal-fired units and meteorological characteristic data affecting load changes, preprocess the historical operating data and meteorological characteristic data, generate model input characteristic data, and divide the model input characteristic data into training set, validation set and test set; The model building module is used to construct the DTA-Autoformer prediction model. It inputs the model's feature data into the DTA-Autoformer prediction model, projects the model's feature data to a unified model dimension, and generates a high-dimensional feature sequence. Based on the high-dimensional feature sequence, it generates a bidirectional time-series representation. Based on the bidirectional time-series representation, it generates dynamic attention weights for each time step. The dynamic attention weights are weighted with the high-dimensional feature sequence and added to the high-dimensional feature sequence through a residual connection domain to obtain a fused feature sequence. The fused feature sequence is then autocorrelated to generate coded features. After cross-correlation interaction using the coded features, trend and periodic terms are extracted and fused through sequence decomposition to generate a fused feature of trend and periodic terms. Finally, the fused feature of trend and periodic terms is mapped to the target dimension to generate a normalized load prediction sequence. The model training module is used to iteratively train the input DTA-Autoformer prediction model using the training set and validation set, and evaluate the prediction performance of the DTA-Autoformer prediction model using the test set to obtain the trained DTA-Autoformer prediction model. The prediction output module is used to input the collected historical operating data of coal-fired units and meteorological characteristic data affecting load changes into the trained DTA-Autoformer prediction model, perform inverse normalization processing on the output normalized load prediction sequence, and generate short-term load prediction results for coal-fired units.

[0014] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a short-term load forecasting method for coal-fired power units.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a short-term load forecasting method for coal-fired power units.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: This invention proposes a short-term load forecasting method for coal-fired power units. By constructing a DTA-Autoformer forecasting model, it organically combines a dynamic temporal attention module with the Autoformer architecture, effectively solving the technical problem that traditional coal-fired power unit load forecasting methods struggle to simultaneously consider long-term trends and localized sudden fluctuations. The dynamic temporal attention module extracts bidirectional contextual information of the sequence through a bidirectional gated cyclic unit and adaptively generates weights for each time step using an attention scoring mechanism. This enhances the model's ability to perceive periods of rapid load change and overcomes the shortcomings of traditional Autoformers, which overemphasize global periodicity and are slow to respond to non-periodic fluctuations. The Autoformer's sequence decomposition and autocorrelation mechanisms can efficiently separate the trend and periodic components of the load sequence, accurately capturing long-distance temporal dependencies. This invention integrates coal-fired power unit operating data, meteorological characteristics, and temporal characteristics for multi-dimensional modeling. Combined with a standardized data preprocessing workflow, it significantly improves the accuracy and stability of short-term load forecasting for coal-fired power units, providing reliable data support for optimized unit operation and grid dispatching. Attached Figure Description

[0017] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely schematic to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. In the drawings: Figure 1 This is a flowchart of the short-term load prediction method for coal-fired power units according to the present invention.

[0018] Figure 2 This is a simplified structural diagram of the short-term load forecasting system for coal-fired power units according to the present invention.

[0019] Figure 3 This is a diagram of the electronic equipment used in the short-term load forecasting method for coal-fired power units according to the present invention.

[0020] Figure 4 This is an overall flowchart of the short-term load forecasting method for coal-fired power units in this embodiment of the invention.

[0021] Figure 5 This is a Pearson correlation coefficient diagram of the characteristic variables and load in an embodiment of the present invention.

[0022] Figure 6 This is a schematic diagram of the Autoformer structure in an embodiment of the present invention.

[0023] Figure 7 This is an overall architecture diagram of the DTA-Autoformer prediction model in this embodiment of the invention.

[0024] Figure 8 This is a curve comparing the predicted value and the actual value in an embodiment of the present invention.

[0025] Figure 9 This is a comparison chart of prediction results from different models in this embodiment of the invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] Dynamic Temporal Attention (DTA); The Autoformer model is an efficient variant of the Transformer designed specifically for long-term time series forecasting; MAE (Mean Absolute Error); RMSE (Root Mean Squared Error); MAPE (Mean Absolute Percentage Error); R² (R-squared, coefficient of determination).

[0028] Example 1 See Figure 1 The method for short-term load forecasting of coal-fired power units includes the following steps: Historical operating data of coal-fired power units and meteorological characteristic data affecting load changes are obtained. The historical operating data and meteorological characteristic data are preprocessed to generate model input characteristic data, and the model input characteristic data are divided into training set, validation set and test set. A DTA-Autoformer prediction model is constructed by inputting the model's input feature data into the model and projecting the data onto a unified model dimension to generate a high-dimensional feature sequence. A bidirectional time-series representation is generated based on this high-dimensional feature sequence. Dynamic attention weights for each time step are generated based on this bidirectional time-series representation. These dynamic attention weights are weighted with the high-dimensional feature sequence and then added to the high-dimensional feature sequence through a residual connection domain to obtain a fused feature sequence. The fused feature sequence is then autocorrelated to generate coded features. After cross-correlation interaction using these coded features, trend and periodic terms are extracted and fused through sequence decomposition to generate a fused feature of trend and periodic terms. Finally, this fused feature of trend and periodic terms is mapped to the target dimension to generate a normalized load prediction sequence. The DTA-Autoformer prediction model is iteratively trained using the training and validation sets, and its prediction performance is evaluated using the test set to obtain the trained DTA-Autoformer prediction model. The historical operating data of coal-fired power units and meteorological characteristic data affecting load changes are input into the trained DTA-Autoformer prediction model. The output normalized load prediction sequence is then denormalized to generate short-term load prediction results for coal-fired power units.

[0029] Specifically, such as Figure 4 As shown, this embodiment of the invention provides a short-term load forecasting method for coal-fired power units, based on dynamic time-series attention Autoformer, including the following steps: S1: Data Acquisition: Collect historical operating data sequences of coal-fired power units and meteorological characteristic data that affect load changes.

[0030] In this embodiment, the data comes from a power plant's SIS (System-In-Service) system, with a collection period spanning one year and a sampling interval of 15 minutes. The unit operating parameters include, but are not limited to: generator active power (MW), total coal consumption (t / h), heat rate (kJ / kWh), boiler efficiency (%), plant power consumption (%), power generation coal consumption g / (kW·h), main steam temperature (°C), reheat steam temperature (°C), and sulfur dioxide concentration (mg / m³). 3 ), nitrogen oxide concentration (mg / m 3 The meteorological data include: HGI, coal feed rate (t / h) of the coal mill AF, coal mill outlet temperature (°C), and total flow rate of desuperheating water in the reheater (t / h); the meteorological characteristic data include: ambient temperature (°C), relative humidity (%), rainfall (mm), and atmospheric pressure (hPa).

[0031] It should be noted that the collected operational data should be data under normal operating conditions of the unit, excluding data records under abnormal operating conditions such as start-up, shutdown, and fault maintenance.

[0032] S2: Preprocess the historical operational data sequence. This preprocessing includes outlier detection and cleaning, missing value imputation, normalization, temporal feature encoding, and feature selection based on the Pearson correlation coefficient. The data processing in step 2 includes: S2.1: Outlier Detection and Cleaning: An outlier detection and cleaning method based on the interquartile range (IQR) is used. The specific process is as follows: Calculate the first quartile Q1 and the third quartile Q3 for each variable to obtain the interquartile range (IQR) = Q3 - Q1. Set the outlier detection boundary as [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR]. Replace outliers below the lower boundary with Q1 - 1.5 × IQR, and replace outliers above the upper boundary with Q3 + 1.5 × IQR. The advantage of this method is that it does not rely on data distribution assumptions, is applicable to skewed data, and is computationally simple and robust.

[0033] S2.2: Missing Value Imputation: Missing value imputation uses linear interpolation. For a time point t with missing values, linear interpolation is performed using the adjacent valid data points before and after it. The formula is as follows:

[0034] in, To fill in the time point t Valid data points x t-k For time points t The most recent valid data point, xt+m For time points t The most recent valid data point after that, k and m These represent the corresponding time intervals.

[0035] S2.3: Pearson Correlation Analysis and Feature Selection: Calculating Each Feature Variable X With target load variable Y The Pearson correlation coefficient r between them is calculated using the following formula:

[0036] In the formula: X i and Y i They represent the first i indivual X and Y The observed values, and They are respectively X and Y The mean.

[0037] When |r|>0.5, the variable is considered to have a strong correlation with the load and is retained as a model input feature. For example... Figure 5 The image shows the Pearson correlation coefficient analysis.

[0038] Feature variables with an absolute correlation coefficient greater than 0.5 were selected as inputs, and meteorological factors were combined. Table 1 below shows the model input variables.

[0039] Table 1 Model Input Variables

[0040] S2.4: Time feature encoding. Hour features H∈{0,1,...,23} and weekday features W∈{0,1,...,6} are extracted from the original timestamp. These time features are then converted using a sine-cosine periodic encoding method, as shown in the following formula:

[0041]

[0042] in, W sin and W cos The sine and cosine coding components for the weekday feature, H sin and H cos The sine and cosine encoded components of the hourly features are used to preserve the cyclical physical characteristics of the time series.

[0043] The encoded 4-dimensional time features are concatenated with the original running parameter features to form the complete input feature vector of the model.

[0044] S2.5: Normalization is performed using the min-max normalization method to normalize the feature data to the range [0,1]. The normalization formula is as follows:

[0045] x ' i Indicates the first i The normalized values ​​of the data. x i Indicates the first i The original data values ​​of each data sample. x max and x min These represent the maximum and minimum values ​​of the original data, respectively.

[0046] It is important to note that the normalization parameter ( x max and x min The normalization parameters are calculated only from the training set, and the validation and test sets are processed using the same normalization parameters to avoid data leakage.

[0047] S2.6: Dataset partitioning. The preprocessed data is divided into training, validation, and test sets in chronological order, with a ratio of 7:1.5:1.5. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and early stopping detection, and the test set is used for final model performance evaluation.

[0048] S3: As Figure 6 and Figure 7 As shown, a DTA-Autoformer prediction model is constructed, which sequentially includes an input projection layer, a dynamic temporal attention module, an Autoformer encoder, an Autoformer decoder, and an output prediction layer.

[0049] In step S3, the construction and calculation process of the dynamic temporal attention module includes: S3.1 Linear Mapping: The selected feature variables are used to construct a multi-dimensional input vector, which is then projected onto a unified model dimension through a linear mapping layer.

[0050]

[0051]

[0052] Where t=1, 2, … L, L is the length of the input sequence, i.e. the number of historical time steps used by the model; u t (load) , u t (humidity) , u t (coal) These represent characteristic values ​​such as load, humidity, and total coal quantity at time t; Indicates that the input feature vector belongs to d 3D real space; U The multidimensional feature input sequence before projection; d B is the number of feature variables; B (batch size) is the batch size. This is the high-dimensional feature vector after mapping; W p The representation of input features after linear mapping and normalization; d p This refers to the random inactivation probability introduced in the linear mapping layer to prevent model overfitting; LayerNorm represents the model encoding dimension.

[0053] S3.2: The mapped sequence input bidirectional gated recurrent unit (BiGRU) extracts forward and backward context information simultaneously in the time dimension:

[0054] in, These represent the hidden states of the Gated Recurrent Unit (GRU) in the forward and reverse directions, respectively. The bidirectional temporal representation vector, which is the concatenation of the forward and backward hidden states, belongs to 2. h 3D real space, h is the feature dimension of the unidirectional hidden layer of the bidirectional gated loop unit.

[0055] S3.3: Obtain the raw attention score for each time step through a two-layer fully connected network:

[0056] in, s t Indicates the first t Original attention score for each time step; W 1, W2. Parameters of fully connected Layer 1 and Layer 2 networks. b 1. b 2 represents the bias term parameters for the first and second layer fully connected networks, respectively.

[0057] S3.4: Weight Normalization: The unnormalized attention score S is normalized using Softmax (normalized exponential function) over time to generate a dynamic time weight vector.

[0058] in, α t Let be the attention weight at the i-th time step; k It is the index variable of the summation symbol. s k It is the first k The raw attention score for each time step. The denominator of the formula is the sum of all time steps within the time window (from 1 to ...). L The exponential scores of the data are locally accumulated to determine the current time step. t The relative weight percentage in the entire sequence.

[0059] S3.5: Feature Enhancement and Residual Fusion: The dynamic time weight vector α is combined with the high-dimensional feature sequence. Z Element-wise multiplication is performed to obtain weighted features, and the weighted features are then multiplied by the high-dimensional feature sequence. Z The enhanced feature sequence is obtained by adding the residual connections. X .

[0060]

[0061] in, It is the feature representation after attention weighting. This represents the enhanced features after residual fusion.

[0062] S3.6: Enhanced sequence X As input to the Autoformer encoder:

[0063] S3.7: Autoformer encoders consist of several stacked encoder layers, such as... Figure 6 As shown, each encoder layer contains a Series Decomposition Block and an Auto-Correlation Mechanism.

[0064] The Series Decomposition Block uses a moving average filter to decompose the input sequence into a trend term and a period term.

[0065] in, X Let be the latent variables to be decomposed. X t , X s ∈R L×d These represent the trend term and the period term, respectively, where d is the model encoding dimension. The above formula can be written as: X t , X s =SeriesDecomp ( X ), AvgPool ( · ) represents a one-dimensional average pooling operation. Padding ( · ) indicates boundary padding to keep the sequence length unchanged.

[0066] The purpose of sequence decomposition is to eliminate the interference of trend terms, so that subsequent autocorrelation mechanisms can more effectively capture periodic features.

[0067] The calculation process of the autocorrelation mechanism is as follows: a) Periodic dependency detection: Calculate the autocorrelation function of the query sequence Q and the key sequence K using FFT (Fast Fourier Transform). :

[0068] Where FFT(·) and IFFT(·) represent the Fast Fourier Transform and Inverse Fourier Transform, respectively. Indicates complex conjugation. τ Indicates a time delay.

[0069] b) Delay selection: Select Top-based values ​​based on autocorrelation. c (forward c The highest value, c Let τ1, τ2, ..., τ be the number of significant delays selected by the hyperparameter. c .

[0070] c) Time-delay aggregation: Roll-align the value sequence V by a selected delay and aggregate it based on autocorrelation values. :

[0071] in,Q , K , V These are the query sequence, key sequence, and value sequence, respectively. Roll( V , τ p ) indicates that the sequence V Circular shift τ p Position, subscript p ∈[1, c [Indicates the selected] c A significant delay is used for summation and traversal.

[0072] The autocorrelation mechanism has a computational complexity of O(Llog L), which is more efficient than the standard attention mechanism's computational complexity of O(L²), and it can effectively capture periodic dependencies in long sequences.

[0073] S3.8: The Autoformer decoder contains multiple stacked decoder layers, such as... Figure 6 As shown, each decoder layer contains an autocorrelation mechanism module, an encoder-decoder interactive autocorrelation module, and a sequence decomposition module.

[0074] The decoder's specific workflow is as follows: (1) Input initialization Using the latter half of the input sequence (of length L / 2, where L is the length of the model input feature sequence mentioned above) as the start token, and concatenating it with a zero vector placeholder (of length P, where P is the prediction length) as the input to the prediction part, the decoder input is constructed:

[0075] in, This represents extracting the latter half of the input historical sequence X (from the L / 2th time step to the last time step L). Concat represents the matrix concatenation operation (joining different feature matrices end-to-end in the time dimension). The sequence decomposition module initializes the decoder input as a periodic term input. X s and trend input X t .

[0076] (2) Hierarchical processing and accumulation In each decoder layer, the periodic term input undergoes the following processing in sequence: a) Autocorrelation mechanism module: captures the periodic dependencies of the decoder input itself.

[0077] b) Encoder-decoder interaction autocorrelation module: Using decoder features as queries and encoder outputs as keys, information interaction is established between the encoder and decoder.

[0078] c) Sequence decomposition module: Separates trend components from processed features.

[0079] (3) Trend convergence The trend components extracted from each decoder layer are summed with the initial trend term input to obtain the final trend prediction part:

[0080] Among them, T m For the first m Trend components extracted from each decoder layer.

[0081] (4) Final output The periodic term prediction part S output from the last decoder layer final The trend prediction part T obtained by summation final Add them together to get the final prediction:

[0082] S3.9: Output Prediction Layer The output prediction layer maps the decoder output to the target dimension through a linear transformation, resulting in a normalized load prediction sequence.

[0083] in, pred ∈R B×L×9 This is the predicted load sequence.

[0084] S4: Input the training set into the DTA-Autoformer prediction model for iterative training, use the loss function to calculate the error between the predicted value and the true value, use the optimizer to update the model parameters based on backpropagation of the error, and perform early stopping judgment based on the validation set until the model converges and saves the optimal model parameters.

[0085] S4.1: Using Mean Squared Error (MSE) as the loss function, calculate the error between the predicted and actual values:

[0086] Where n is the number of samples. j For the first j The predicted value for each sample, y j For the first j The true value of each sample.

[0087] The MSE loss function is sensitive to large errors and can effectively penalize large prediction biases, guiding the model to optimize prediction accuracy.

[0088] S4.2: The optimizer uses the AdamW (Adaptive Moment Estimation with Decoupled Weight Decay) optimizer for parameter updates. AdamW introduces a decoupled weight decay regularization term on top of the standard Adam (Adaptive Moment Estimation) optimizer, which can effectively prevent model overfitting.

[0089] AdamW's parameter update formula is:

[0090]

[0091]

[0092]

[0093]

[0094] in, m step This represents the first moment estimate of the gradient. v step The second-order primitive moment estimate of the gradient is given. θ step These are the parameters of the model at the step-th iteration. This is the first-order moment estimate after bias correction. This is the second-order primitive moment estimate after bias correction. g step For gradient, β 1 and β 2 is the momentum coefficient. η For learning rate, ε For numerically stable terms, λ This is the weight decay coefficient. Indicates the t-th The first moment estimate of the gradient in one iteration. express β 1 raised to the power of the current number of iterations completed. express β 2 is the power of the number of iterations that have been completed so far.

[0095] In this embodiment, the hyperparameters are set as follows: initial learning rate η = 0.002, β1 = 0.9, β2 = 0.999, ε = 10. -8 The weight decay coefficient λ=0.01, batch size=64, model encoding dimension dmodel=128, number of Autoformer encoder layers N=3, number of decoder layers M=4, and BiGRU hidden layer dimension=256.

[0096] S4.3: To prevent model overfitting, an early stopping strategy based on the validation set is adopted. The specific process is as follows: After each training round, the validation loss Lval is calculated using the validation set: if Lval is better than the historical best Lbest (Best Validation Loss), then Lbest is updated, the optimal parameters are saved, and the waiting counter is reset; otherwise, the counter is incremented by 1. When the counter reaches the preset patience threshold of 60, training is stopped and the optimal parameters are loaded.

[0097] S5: Obtain the normalized load forecast sequence, and perform inverse normalization on the normalized load forecast sequence to obtain the short-term load forecast values ​​for coal-fired units. The inverse normalization formula is:

[0098] Where y' is the normalized forecast value, and y is the inverse normalized actual load forecast value. max and y min The maximum and minimum values ​​of the training load data.

[0099] Experimental verification: To verify the effectiveness of the method of this invention, experiments were conducted using actual operating data from a 600MW coal-fired power unit. Experimental data: The actual load data was selected as the dataset, with a total of 2017 sets of data. The sampling interval was 15 minutes. The training set accounted for 70% of the total data, the validation set accounted for 15% of the total data, and the test set accounted for 15% of the total data. The model was then experimentally validated.

[0100] The following four indicators are used to evaluate predictive performance: (1) MAE (Mean Absolute Error):

[0101] (2) RMSE (Root Mean Squared Error):

[0102] (3) MAPE (Mean Absolute Percentage Error):

[0103] (4) R² (R-squared, coefficient of determination):

[0104] in, This represents the true observation value of the i-th sample. This represents the model prediction value for the i-th sample. This represents the arithmetic mean of the true values ​​of all test samples.

[0105] Comparative experiment: The proposed DTA-Autoformer method is compared with the following benchmark models: LSTM (Long Short-Term Memory), BiLSTM (Bidirectional Long Short-Term Memory), Transformer, CNN (Convolutional Neural Network), and Autoformer. See Tables 2 and 3. Figure 8 , Figure 9 Experimental results show that the DTA-Autoformer method proposed in this invention outperforms the comparative model in all evaluation indicators. Compared with the standard Autoformer, RMSE is reduced by approximately 19.15%, MAE by approximately 16.47%, and R² is improved by approximately 2.4%. Especially during periods of rapid load change, the DTA-Autoformer can more accurately track load change trends, and the prediction lag is significantly reduced.

[0106] Table 2 Model Indicator Evaluation

[0107] Table 3 Evaluation Indicators for Comparison of Different Models

[0108] The DTA-Autoformer model proposed in this invention innovatively embeds a dynamic temporal attention module into the standard Autoformer architecture. This module uses bidirectional gated recurrent units to capture the bidirectional contextual dependencies of the sequence and adaptively generates time step weights through an attention scoring network, which effectively enhances the model's ability to perceive local dynamic changes and aperiodic fluctuations, and overcomes the problem of slow response of traditional Autoformer models to sudden load changes.

[0109] This invention transforms the time features using a sine-cosine periodic encoding method, concatenates the encoded time features with the original operating parameter features, and incorporates meteorological factors as multi-dimensional input features for the model. Combined with a Dynamic Temporal Attention (DTA) module, this significantly enhances the model's ability to utilize multi-dimensional information, which is the core advantage of this invention over traditional univariate decomposition prediction methods.

[0110] This invention integrates the sequence decomposition mechanism and autocorrelation mechanism of Autoformer, which can effectively separate the trend term and periodic term in the load sequence, capture long-term dependencies, and combine the local feature extraction capability of the dynamic time series attention module to achieve high-precision prediction of short-term load of coal-fired units.

[0111] Example 2 See Figure 2 The short-term load forecasting system for coal-fired power units includes: The data processing module is used to acquire historical operating data of coal-fired units and meteorological characteristic data affecting load changes, preprocess the historical operating data and meteorological characteristic data, generate model input characteristic data, and divide the model input characteristic data into training set, validation set and test set; The model building module is used to construct the DTA-Autoformer prediction model. It inputs the model's feature data into the DTA-Autoformer prediction model, projects the model's feature data to a unified model dimension, and generates a high-dimensional feature sequence. Based on the high-dimensional feature sequence, it generates a bidirectional time-series representation. Based on the bidirectional time-series representation, it generates dynamic attention weights for each time step. The dynamic attention weights are weighted with the high-dimensional feature sequence and added to the high-dimensional feature sequence through a residual connection domain to obtain a fused feature sequence. The fused feature sequence is then autocorrelated to generate coded features. After cross-correlation interaction using the coded features, trend and periodic terms are extracted and fused through sequence decomposition to generate a fused feature of trend and periodic terms. Finally, the fused feature of trend and periodic terms is mapped to the target dimension to generate a normalized load prediction sequence. The model training module is used to iteratively train the DTA-Autoformer prediction model using the training set and validation set, and evaluate the prediction performance of the DTA-Autoformer prediction model using the test set to obtain the trained DTA-Autoformer prediction model. The prediction output module is used to input the collected historical operating data of coal-fired units and meteorological characteristic data affecting load changes into the trained DTA-Autoformer prediction model, perform inverse normalization processing on the output normalized load prediction sequence, and generate short-term load prediction results for coal-fired units.

[0112] Example 3 See Figure 3 An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a short-term load forecasting method for coal-fired power units.

[0113] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for short-term load forecasting of coal-fired power units.

[0114] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, read-only optical discs, optical storage, etc.) containing computer-usable program code.

[0115] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for short-term load forecasting of coal-fired power units, characterized in that, Includes the following steps: Historical operating data of coal-fired power units and meteorological characteristic data affecting load changes are acquired. The historical operating data includes generator active power, total coal consumption, heat rate, boiler efficiency, plant power consumption, power generation coal consumption, main steam temperature, reheat steam temperature, sulfur dioxide concentration, nitrogen oxide concentration, HGI, coal consumption of pulverizer AF, pulverizer outlet temperature, and total flow rate of reheater desuperheating water. The meteorological characteristic data includes ambient temperature, relative humidity, rainfall, and atmospheric pressure. The historical operating data and the meteorological characteristic data are preprocessed to generate model input characteristic data, which are then divided into training set, validation set, and test set. A DTA-Autoformer prediction model is constructed by inputting the model's input feature data into the DTA-Autoformer prediction model. The DTA-Autoformer prediction model sequentially includes an input projection layer, a dynamic temporal attention module, an Autoformer encoder, an Autoformer decoder, and an output prediction layer. The input projection layer projects the model's input feature data onto a unified model dimension, generating a high-dimensional feature sequence. The dynamic temporal attention module includes a bidirectional gated recurrent unit, an attention scoring unit, and a residual fusion unit. The bidirectional gated recurrent unit receives the high-dimensional feature sequence and outputs a bidirectional temporal representation. The attention scoring unit receives the bidirectional temporal representation through a two-layer fully connected network, generates raw attention scores for each time step, and performs Softmax normalization on the raw attention scores in the temporal dimension to generate dynamic attention weights for each time step. The residual fusion unit multiplies the dynamic attention weights element-wise with the high-dimensional feature sequence to obtain weighted features. It then adds the weighted features to the high-dimensional feature sequence via a residual connection, outputting the fused feature sequence to the Autoformer encoder. The fused feature sequence undergoes autocorrelation encoding to generate encoded features. After cross-correlation interaction using the encoded features, trend and periodic terms are extracted and fused through sequence decomposition to generate a fused feature of trend and periodic terms. This fused feature is mapped to the target dimension to generate a normalized load prediction sequence. The Autoformer encoder encodes the fused feature sequence. The Autoformer decoder receives the output of the Autoformer encoder and decodes it. The output prediction layer maps the output of the Autoformer decoder to the target dimension to generate a normalized load prediction sequence. The DTA-Autoformer prediction model is iteratively trained using the training set and the validation set, and the prediction performance of the DTA-Autoformer prediction model is evaluated using the test set to obtain the trained DTA-Autoformer prediction model. The historical operating data of the coal-fired power units and the meteorological characteristic data affecting load changes are input into the trained DTA-Autoformer prediction model. The normalized load prediction sequence output is then denormalized to generate short-term load prediction results for the coal-fired power units.

2. The short-term load forecasting method for coal-fired power units according to claim 1, characterized in that, The preprocessing of the historical operational data and the meteorological characteristic data includes: The historical operational data and meteorological feature data are subjected to outlier detection and cleaning, missing value imputation, feature screening based on Pearson correlation coefficient, time feature encoding, and normalization processing to generate the model input feature data.

3. The short-term load forecasting method for coal-fired power units according to claim 1, characterized in that, The model input feature data includes: Features selected from the historical operational data whose absolute value of the load correlation coefficient is greater than a preset limit; Temperature, humidity, and rainfall obtained from the meteorological characteristic data; And hourly and weekday features extracted from the original timestamps and periodically encoded using sine and cosine.

4. The short-term load forecasting method for coal-fired power units according to claim 1, characterized in that, The Autoformer encoder consists of multiple stacked encoder layers, each of which includes an autocorrelation mechanism module and a sequence decomposition module in sequence. The sequence decomposition module uses a moving average filter to decompose the input sequence into a trend term and a period term. The autocorrelation mechanism module calculates the autocorrelation function of the sequence through fast Fourier transform, selects the top-c significant delays, performs rolling alignment of the value sequence according to the selected delays, and aggregates them based on the autocorrelation values.

5. The short-term load forecasting method for coal-fired power units according to claim 1, characterized in that, The Autoformer decoder consists of multiple stacked decoder layers, each of which sequentially includes a first autocorrelation mechanism module, an encoder-decoder interactive autocorrelation mechanism module, and a sequence decomposition module. The Autoformer decoder uses the latter half of the input sequence as a starting marker, concatenates zero vector placeholders to construct the decoder input, and initializes the decoder input as periodic term input and trend term input through the sequence decomposition module; In each decoder layer, the periodic term input is processed sequentially by the first autocorrelation mechanism module, the encoder-decoder interactive autocorrelation mechanism module, and the sequence decomposition module. The separated trend component is accumulated with the initial trend term input, and finally the periodic term prediction part is added with the accumulated trend prediction part to obtain the decoding result.

6. The method for short-term load forecasting of coal-fired power units according to claim 1, characterized in that, The iterative training of the DTA-Autoformer prediction model using the training set and the validation set includes: The mean squared error is used as the loss function to calculate the error between the predicted value and the true value. The AdamW optimizer is used to update the model parameters based on the backpropagation of the error; An early stopping strategy is implemented based on the prediction error of the validation set. Training is stopped when the validation loss fails to improve for a preset number of consecutive rounds, and the optimal model parameters are saved.

7. A short-term load forecasting system for coal-fired power units, using the short-term load forecasting method for coal-fired power units as described in any one of claims 1-6, characterized in that, include: The data processing module is used to acquire historical operating data of coal-fired power units and meteorological characteristic data affecting load changes, preprocess the historical operating data and the meteorological characteristic data to generate model input characteristic data, and divide the model input characteristic data into training set, validation set and test set; The model building module is used to construct a DTA-Autoformer prediction model. The model input feature data is input into the DTA-Autoformer prediction model, and the input feature data is projected onto a unified model dimension to generate a high-dimensional feature sequence. A bidirectional temporal representation is generated based on the high-dimensional feature sequence. Dynamic attention weights for each time step are generated based on the bidirectional temporal representation. The dynamic attention weights are weighted with the high-dimensional feature sequence and added to the high-dimensional feature sequence through a residual connection domain to obtain a fused feature sequence. Autocorrelation encoding is performed on the fused feature sequence to generate encoded features. After cross-correlation interaction using the encoded features, trend terms and periodic terms are extracted and fused through sequence decomposition to generate a fused feature of trend and periodic terms. The trend term and periodic term are fused and mapped to the target dimension to generate a normalized load prediction sequence. The model training module is used to iteratively train the DTA-Autoformer prediction model using the training set and the validation set, and evaluate the prediction performance of the DTA-Autoformer prediction model using the test set to obtain the trained DTA-Autoformer prediction model. The prediction output module is used to input the collected historical operating data of coal-fired units and meteorological characteristic data affecting load changes into the trained DTA-Autoformer prediction model, and to perform inverse normalization processing on the output normalized load prediction sequence to generate short-term load prediction results for coal-fired units.

8. An electronic device, characterized in that, The method includes 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 the short-term load forecasting method for coal-fired power units as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the short-term load forecasting method for coal-fired power units as described in any one of claims 1-6.

Images