Power consumption prediction method and system for silicon-based enterprise carbon footprint checking

Abstract

The invention discloses a silicon-based enterprise carbon footprint verification-oriented power consumption prediction method and system, and belongs to the technical field of silicon-based industry power consumption power prediction, and the method comprises the steps: obtaining and preprocessing silicon-based enterprise power consumption power time sequence data; constructing a power consumption prediction model for silicon-based enterprise carbon footprint checking; training a power consumption prediction model based on the preprocessed power consumption power time sequence data of the silicon-based enterprise; and performing prediction based on the trained power consumption prediction model to obtain a power consumption prediction result. The method effectively solves the technical bottleneck that a traditional single-scale model is difficult to consider long-term trend and short-term fluctuation at the same time, the characterization capability of the model for the multi-scale time characteristic of the power consumption power data of the silicon-based enterprise is remarkably improved, and then the prediction accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of power consumption prediction technology in the silicon-based industry, and in particular to a power consumption prediction method and system for carbon footprint verification of silicon-based enterprises. Background Technology

[0002] The silicon-based industry, as a key core material industry for strategic emerging industries such as electronic information, new energy, and energy conservation and environmental protection, is characterized by high energy consumption and high emissions. By calculating carbon emissions, enterprises can more clearly understand their shortcomings in the production process, thereby increasing technological innovation and industrial upgrading, improving product quality and added value, which is of great practical significance for achieving the strategic goals of regional carbon emission reduction.

[0003] The power consumption of silicon-based enterprises is characterized by intermittency, volatility, and uncertainty, exhibiting complex nonlinear time-varying properties. This inherent randomness makes it difficult to accurately predict and effectively control the power consumption of silicon-based enterprises. Power consumption prediction for silicon-based enterprises, as a core supporting technology, plays a crucial role in sensing changes in power consumption, predicting power consumption capabilities, and driving scheduling strategies. It is of paramount importance in improving power dispatch efficiency and ensuring power supply quality.

[0004] However, existing research still faces key challenges in assessing the power consumption of silicon-based enterprises. First, most methods model only a single timescale, making it difficult to simultaneously capture multi-level features such as long-term trends, cyclical patterns, and short-term fluctuations. This limits the model's expressive power and results in insufficient prediction accuracy in complex scenarios requiring consideration of multiple time granularities. Second, although the Transformer model has been introduced to capture long-term dependencies, its global attention mechanism suffers from two problems: high computational complexity and susceptibility to noise, weakening its ability to capture key local patterns and abrupt changes. Furthermore, existing methods lack effective cross-scale interaction and adaptive weight adjustment mechanisms when fusing multi-scale features, relying heavily on simple concatenation or fixed weight strategies, which limits the effectiveness of multi-scale collaborative modeling.

[0005] Therefore, there is an urgent need for a silicon-based enterprise power consumption method that can effectively capture multi-scale temporal features, balance local and global dependency modeling, and achieve adaptive fusion of cross-scale information. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this application provides a method and system for predicting electricity consumption for carbon footprint verification of silicon-based enterprises. This method solves the problems of existing technologies, such as difficulty in simultaneously capturing the multi-scale time features contained in the electricity consumption data of silicon-based enterprises, the neglect of the heterogeneity of local and global dependencies in the electricity consumption data of silicon-based enterprises by existing Transformer models, and the failure to consider the scenario and data characteristics in order to determine the contribution of features at each scale.

[0007] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: First aspect: This application provides a method for predicting electricity consumption for carbon footprint verification of silicon-based enterprises, including: S1: Acquire and preprocess time-series power consumption data of silicon-based enterprises; S2: Construct an electricity consumption prediction model for carbon footprint verification of silicon-based enterprises; the electricity consumption prediction model includes a multi-scale decomposition layer, a multi-level parallel encoding and decoding layer, and an adaptive weighted fusion layer; The multi-scale decomposition layer includes a trend component module, a periodic component extraction module, and a proximity component extraction module. The trend component module captures the long-term evolution characteristics of the power consumption time series data of silicon-based enterprises; the periodic component extraction module learns the regularity characteristics of the power consumption time series data of silicon-based enterprises; and the proximity component extraction module captures the short-term fluctuation characteristics of the power consumption time series data of silicon-based enterprises. The multi-level parallel encoding and decoding layer includes a trend encoder, a period encoder, and a proximity encoder. The trend encoder performs deep encoding on long-term evolution features to obtain weekly trend components; the period encoder performs deep encoding on regularity features to obtain daily period components; and the proximity encoder performs deep encoding on short-term fluctuation features to obtain hourly proximity components. The adaptive weighted fusion layer dynamically weights and fuses the three time-scale components using a learnable weight matrix, and automatically adjusts the weight allocation according to the actual contribution of each scale feature to the prediction task. S3: Train an energy consumption prediction model based on preprocessed time series data of power consumption of silicon-based enterprises; S4: Make predictions based on the trained power consumption prediction model to obtain power consumption prediction results.

[0008] Furthermore, the preprocessing includes: The original silicon-based enterprise power consumption time series data was initially screened by batch deletion and correction of erroneous records. Median filtering was used to smooth the remaining outliers. Each data point was replaced with the median of its own value and the adjacent observations. Min-max normalization was used to scale the values. Holiday and weekday time metadata were one-hot encoded.

[0009] Furthermore, the trend encoder, period encoder, and proximity encoder all include a long- and short-distance attention fusion module; The long and short distance attention fusion module includes a short distance attention submodule, a long distance attention submodule, and a multilayer perceptron. The short-range attention submodule uniformly divides the power consumption time series data of silicon-based enterprises into multiple non-overlapping windows of fixed length, and independently calculates the query matrix, key matrix and value matrix within the window. It models the local time dependencies within the window through a scaling dot product attention mechanism and outputs the splicing result of the local time dependencies of all windows. The long-distance attention submodule performs sparse sampling on the power consumption time series data of silicon-based enterprises with a fixed step size and constructs a sparse label set. It then performs global self-attention calculation on the sparse label set. The sparse attention output is restored to the original sequence length through linear interpolation to obtain the long-distance time dependency. The multilayer perceptron performs linear transformation and feature enhancement on the outputs of the short-range attention submodule and the long-range attention submodule.

[0010] Furthermore, the trend component module captures the long-term evolution characteristics of the power consumption time series data of silicon-based enterprises, including: A1: By embedding the input values, a linear transformation is performed on the weekly power consumption time series data of silicon-based enterprises to obtain the content features of the power consumption time series data of silicon-based enterprises. A2: Introduce time embedding to establish a periodic pattern in the weekly power consumption time series data of silicon-based enterprises, and obtain the periodicity of the power consumption time series data of silicon-based enterprises; A3: By injecting the positional relationship of the time step through positional encoding, the positional characteristics of the power consumption time series data of silicon-based enterprises can be obtained; A4: By integrating the content characteristics, periodicity, and location characteristics of the power consumption time series data of silicon-based enterprises, the long-term evolution characteristics of the power consumption time series data of silicon-based enterprises can be obtained.

[0011] Furthermore, the periodic component extraction module learns the regularity characteristics of the power consumption time series data of silicon-based enterprises, including: B1: Encode the daily power consumption time series data of silicon-based enterprises; B2: Utilize multi-head attention mechanisms to extract intraday variation features from encoded data; B3: By introducing a location coding mechanism and combining it with intraday variation characteristics, the regularity characteristics of the power consumption time series data of silicon-based enterprises are obtained.

[0012] Furthermore, the window length of the short-range attention submodule and the sampling step size of the long-range attention submodule include: the window length is equal to the sparse sequence length divided by the sampling step size.

[0013] Furthermore, the adaptive weighted fusion layer also includes a loss optimization module, which uses the mean squared error loss function for end-to-end training to optimize model parameters.

[0014] Furthermore, the step of training the power consumption prediction model based on the preprocessed silicon-based enterprise power consumption time series data includes: C1: The preprocessed power consumption time series data of silicon-based enterprises is divided into weekly trend components, daily periodic components, and hourly proximity components. C2: Train the trend encoder, period encoder, and proximity encoder based on weekly trend components, daily periodic components, and hourly proximity components. C3: Based on the trained trend encoder, period encoder and proximity encoder, capture the long-term evolution characteristics, regularity characteristics and short-term fluctuation characteristics of the power consumption time series data of silicon-based enterprises, and obtain weekly trend components, daily period components and hourly proximity components. C4: Adaptive weighted fusion of the weekly trend component, daily periodic component, and hourly proximity component of the power consumption time series data of silicon-based enterprises is performed using a learnable weight matrix to obtain power consumption prediction results. C5: Based on the power consumption prediction results, the parameters of the power consumption prediction model are optimized using the mean square error loss function.

[0015] The second aspect: This application provides an electricity consumption prediction system for carbon footprint verification of silicon-based enterprises, including: a data acquisition module, a trend component extraction module, a periodic component extraction module and a proximity component extraction module set in parallel, a trend encoder, a periodic encoder and a proximity encoder set in parallel, and an adaptive weighted fusion module. The data acquisition module acquires and preprocesses time-series data on power consumption of silicon-based enterprises; The trend component module captures the long-term evolution characteristics of the power consumption time series data of silicon-based enterprises; the periodic component extraction module learns the regularity characteristics of the power consumption time series data of silicon-based enterprises; and the proximity component extraction module captures the short-term fluctuation characteristics of the power consumption time series data of silicon-based enterprises. The trend encoder performs deep encoding on long-term evolution features to obtain weekly trend components; the period encoder performs deep encoding on regularity features to obtain daily period components; and the proximity encoder performs deep encoding on short-term fluctuation features to obtain hourly proximity components. The adaptive weighted fusion module dynamically weights and fuses the three time-scale components using a learnable weight matrix to obtain the power consumption prediction result.

[0016] Furthermore, the power consumption prediction system for carbon footprint verification of silicon-based enterprises also includes a visualization analysis module, which displays a comparison curve between the prediction results and the actual values.

[0017] The beneficial effects of this application are: This application provides a method and system for predicting electricity consumption for carbon footprint verification of silicon-based enterprises. By constructing three parallel processing paths to capture weekly trends, daily cycles, and hourly proximity features, it effectively solves the technical bottleneck of traditional single-scale models that cannot simultaneously take into account long-term trends and short-term fluctuations, significantly improving the model's ability to represent the multi-scale temporal characteristics of power consumption data of silicon-based enterprises. Furthermore, the adaptive weighted fusion strategy automatically adjusts the contribution weights of features at different time scales through end-to-end learning, avoiding the subjectivity and limitations of traditional fixed-weight fusion methods. This achieves dynamic collaborative optimization of cross-scale information, enhancing the model's generalization ability and adaptability in different prediction scenarios. In addition, the encoder innovatively combines local window attention and global sparse attention through a long-short distance attention fusion mechanism. This ensures both sensitive response to local mutations and effective capture of global evolution patterns, overcoming the inherent defects of high computational complexity and excessive smoothing in standard Transformer global attention, and significantly improving prediction accuracy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart illustrating a method for predicting electricity consumption for carbon footprint verification of silicon-based enterprises, provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of an electricity consumption prediction model for carbon footprint verification of silicon-based enterprises, provided as an embodiment of this application.

[0021] Figure 3 A comparison chart of power consumption results of silicon-based enterprises on the Yamansu dataset provided in this application embodiment.

[0022] Figure 4 A comparison chart of power consumption results of silicon-based enterprises on the Beitashan dataset provided in this application embodiment.

[0023] Figure 5 A comparison chart of power consumption results of silicon-based enterprises on the Shicheng sub-dataset provided in this application embodiment. Detailed Implementation

[0024] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0025] Example 1: Due to the significant multi-scale temporal correlations in the power consumption of silicon-based enterprises, relying solely on single-scale modeling is insufficient to accurately capture its complex dynamic evolution. Existing research still faces key challenges in silicon-based enterprise power consumption data. First, deep learning technology has demonstrated excellent feature extraction and modeling capabilities in silicon-based enterprise power consumption tasks; however, most existing methods focus on modeling at a single time scale, making it difficult to simultaneously capture the multi-scale temporal features inherent in silicon-based enterprise power consumption data. Silicon-based enterprise power consumption sequences simultaneously contain multi-level temporal characteristics such as long-term trends, periodic patterns, and short-term fluctuations. Traditional single-scale modeling methods have two main limitations: first, they struggle to effectively uncover the potential correlations and complementarities between features at different time scales, resulting in limited model expressive power; second, when faced with complex prediction scenarios requiring consideration of multiple time granularities such as hourly, daily, and weekly scales, a single model struggles to simultaneously capture local details and model global trends, leading to insufficient prediction accuracy and generalization ability. Secondly, although Transformer-based time series prediction models have been gradually introduced into the study of power consumption in silicon-based enterprises, aiming to improve modeling performance by capturing long-distance temporal dependencies through self-attention mechanisms, existing Transformer models often overlook the heterogeneity of local and global dependencies in the power consumption data of silicon-based enterprises. The standard Transformer employs a global self-attention mechanism, assigning the same attention weight to all positions in the sequence, which easily leads to two problems: first, the computational complexity increases quadratically with the sequence length, resulting in huge computational costs when predicting long-term sequences; second, global attention introduces a large amount of redundant information and noise, blurring the capture of physically meaningful key temporal patterns, leading to over-smoothing and weakening the model's sensitivity and responsiveness to sudden changes in power consumption in silicon-based enterprises. Furthermore, existing methods lack effective cross-scale information interaction mechanisms and adaptive weight adjustment strategies when dealing with the collaborative modeling of features across multiple time scales, making it difficult to fully explore the potential correlations and complementary information between features at different time scales and achieve true multi-scale collaborative optimization. Traditional methods often employ simple feature concatenation or fixed-weight fusion strategies, failing to dynamically adjust the contribution of features at each scale according to different prediction scenarios and data characteristics, thus limiting further improvements in model performance.

[0026] Based on this, embodiments of this application provide a method for predicting electricity consumption for carbon footprint verification of silicon-based enterprises. This method can be found in [reference needed]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a method for predicting electricity consumption for carbon footprint verification of silicon-based enterprises, as provided in an embodiment of this application. The method includes: S1: Acquire and preprocess the time series data of power consumption of silicon-based enterprises and decompose it into components of three different time scales, including weekly trend components, daily periodic components and hourly proximity components, which correspond to long-term evolution trends, periodic regular patterns and short-term fluctuation characteristics, respectively.

[0027] In one embodiment of this application, the preprocessing includes a preliminary screening of the original silicon-based enterprise power consumption data, which involves batch deletion and correction of erroneous records; a smoothing process for the remaining outliers using median filtering technology; replacing each data point with the median of its own value and the adjacent observations; scaling the values ​​to the range of 0 to 1 using min-max normalization; and one-hot encoding of time metadata such as holidays and days of the week.

[0028] S2: Construct an electricity consumption prediction model for carbon footprint verification of silicon-based enterprises.

[0029] like Figure 2 As shown, the provided power consumption prediction model MCRssFormer for carbon footprint verification methods of silicon-based enterprises includes a multi-scale decomposition layer, a multi-level parallel encoding and decoding layer, and an adaptive weighted fusion layer. The multi-scale decomposition layer includes a trend component module, a periodic component extraction module, and a proximity component extraction module. The trend component extraction module is used to model the long-term evolution trend of power consumption of silicon-based enterprises. This module constructs a weekly trend component from the time series data of power consumption of silicon-based enterprises, with a time window length set to 168 hours, corresponding to 7 days multiplied by 24 hours. Long-term features in the power consumption of silicon-based enterprises often contain multiple complex patterns. To effectively capture these patterns, an end-to-end trend modeling method is designed. Specifically, firstly, the original power sequence is linearly transformed through input value embedding; then, time embedding is introduced to model periodic patterns such as days of the week and months; finally, positional encoding is used to inject the positional relationship of time steps. By summing these three embeddings element-wise, a trend representation integrating content, periodicity, and positional information can be obtained.

[0030] The periodic component extraction module is used to extract the intraday regularity and seasonality features of power consumption data from silicon-based enterprises. This module constructs daily periodic components from the time series data of power consumption of silicon-based enterprises, with a time window length set to 24 hours. The power consumption data of silicon-based enterprises simultaneously exhibits intraday periodicity and seasonal correlation. Regular changes exist between different time periods, and this periodic structure information is crucial for accurately predicting the power consumption of silicon-based enterprises. To extract effective periodic information, the original input for each day is first feature-encoded, and a multi-head attention mechanism is used to process the periodic input to extract intraday variation features. To enhance the global dependency between different dates, a positional encoding mechanism is further introduced. Finally, the periodic features are unified to the dimensions required by the model through an encoder, obtaining the periodic feature representation of each time scale at each time step.

[0031] The proximity component extraction module is used to capture the short-term, rapid fluctuations in the power consumption of silicon-based enterprises. This module constructs hourly proximity components from the time-series power consumption data of silicon-based enterprises, with time window lengths set to 1 to 6 hours. By constructing hourly time windows, rapidly changing dynamic characteristics can be captured, providing crucial information support for real-time power prediction.

[0032] For trend, periodic, and proximity components, a parallel encoding layer is used to construct specialized encoding mechanisms between features at different time scales, fusing deep information from multiple time scales in both local and global dimensions. The parallel encoding layer is responsible for modeling and learning the relationships between feature representations from different time scales, and mainly consists of a trend encoder, a periodic encoder, and a proximity encoder. Each encoder first performs deep encoding on features from different time scales, and then models the dependencies between time scales. In this way, the model can learn various explicit and implicit correlation patterns, such as trend to periodicity, periodicity to proximity, and proximity to trend. When multi-scale synergistic effects exist, the parallel encoding layer enhances cross-scale information transfer. The parallel encoding layer is used to fuse temporal features from the multi-scale decomposition layer across scales, employing an independent encoder structure to model the characteristic differences between different time scales, generating time-scale-specific encoded representations. Furthermore, as... Figure 2 As shown, the encoder has 2 layers, the decoder has 1 layer, the multi-head attention (short-range attention SDA, long-range attention LDA) heads are 8, the model dimension is 128, each layer contains a multi-head attention sub-layer and a feedforward neural network sub-layer (i.e., multilayer perceptron MLP), residual connections and layer normalization (LN) operations are added after each sub-layer, and the hidden layer dimension of the feedforward neural network is set to 4 times the model dimension.

[0033] Understandably, parallel encoding layers are used to construct specialized encoding mechanisms between features at different time scales, fusing deep information from multiple time scales in both local and global dimensions. In doing so, the model can learn various explicit and implicit correlation patterns, such as trend to cycle, cycle to proximity, and proximity to trend. When multi-scale synergistic effects exist, parallel encoding layers will enhance cross-scale information transfer, generating time-scale-specific encoded representations.

[0034] A short-range and long-range attention fusion module is designed in each encoder-decoder module. This module includes a short-range attention submodule, a long-range attention submodule, and a multilayer perceptron. The short-range attention submodule uniformly divides the silicon-based enterprise power consumption time series data into multiple non-overlapping windows of fixed length, and independently calculates the query matrix, key matrix, and value matrix within each window. It models the local temporal dependencies within the window through a scaling dot product attention mechanism and outputs the concatenated result of the local temporal dependencies of all windows. The long-range attention submodule sparsely samples the silicon-based enterprise power consumption time series data with a fixed step size and constructs a sparse label set. It then performs global self-attention computation on the sparse label set. Through the sparse sampling strategy, the computational complexity is reduced from the square of the sequence length to the square of the sampled sequence length. The sparse attention output is restored to the original sequence length through linear interpolation to obtain the long-range temporal dependencies. The alternating stacking of these two modules achieves progressive feature fusion from local to global, effectively solving the problems of high computational complexity and over-smoothing in the standard Transformer. The multilayer perceptron consists of two fully connected layers. The hidden layer dimension is set to 4 times the model dimension. The activation function is Gaussian error linear unit. The final feature representation is generated after residual connection and layer normalization.

[0035] In one embodiment of this application, each encoder-decoder module adds a residual connection after the multi-head attention sublayer. The sublayer input and output are added together, and then layer normalization is performed, effectively alleviating the gradient vanishing problem in deep networks and accelerating model convergence. The trend encoder stacks all time steps of the weekly trend component in chronological order to form a complete input sequence representation. A time-dimensional context representation is extracted using a multi-head attention mechanism. The encoder has 2 layers, the multi-head attention heads are 8, and the model dimension is 128. The periodic encoder and proximity encoder use the same network architecture as the trend encoder, but their parameters are trained independently. They are specifically optimized for daily and hourly time scales, respectively, maintaining the same dimensional structure as the input. Furthermore, the short-range attention submodule and the long-range attention submodule use an alternating stacking structure, applied alternately at different levels of the encoder to achieve progressive fusion of multi-scale features from local to global. The window length of the short-range attention submodule and the sampling step size of the long-range attention submodule satisfy the relationship that the window length equals the sparse sequence length divided by the sampling step size. By adjusting the window length and sampling step size, a trade-off between computational efficiency and modeling accuracy can be achieved.

[0036] The adaptive fusion layer receives the outputs of the trend encoder, the period encoder, and the proximity encoder. It then performs dynamic weighted fusion using a learnable weight matrix to obtain the final prediction result. The weight matrix is ​​jointly optimized with other network parameters through the backpropagation algorithm, and the weight allocation is automatically adjusted according to the actual contribution of features at different time scales to the prediction task.

[0037] In one embodiment of this application, the adaptive fusion layer includes a learnable weight module, a feature fusion module, and a prediction output module. The learnable weight module automatically learns the contribution of features at different time scales to the final prediction; this module dynamically adjusts the weight allocation through end-to-end training. The feature fusion module implements a weighted combination of the trend encoder output, the period encoder output, and the proximity encoder output, improving cross-scale feature alignment capabilities. This process enables the learning model to effectively capture important patterns and information from features at multiple time scales. The prediction output module outputs the final power consumption result for silicon-based enterprises, which comprehensively considers the combined effects of long-term trends, medium-term cycles, and short-term fluctuations. Typically, the output is represented as a prediction sequence over multiple time steps, with the prediction step size flexibly configurable according to actual needs. This framework utilizes the physical and statistical relationships between different time scales of power consumption data from silicon-based enterprises, effectively enabling joint modeling of feature representations and improving prediction accuracy and generalization performance.

[0038] Furthermore, due to the high modeling complexity of multi-scale learning, a large-scale deep learning multi-scale prediction model is created and trained in the system to help accurately predict the power consumption of silicon-based enterprises. This prediction model will be periodically retrained to update its predictive capabilities. Specifically, the adaptive fusion layer also includes a loss optimization module to guide the collaborative learning of multi-scale features during model training. The loss optimization module uses a mean squared error loss function for end-to-end training, calculating the sum of squares of the differences between predicted and true values ​​and dividing by the number of samples, enabling the model to dynamically optimize the modeling quality of each time-scale component at different training stages.

[0039] S3: Train an energy consumption prediction model based on preprocessed time series data of power consumption of silicon-based enterprises.

[0040] In one embodiment of this application, historical power data from power monitoring equipment of silicon-based enterprises is first collected and then converted into multi-scale time feature representations, which are used as input to a learning model. Then, a power consumption prediction model for silicon-based enterprise carbon footprint verification methods is constructed to mine deep temporal semantics and simultaneously model power features at multiple time scales, including weekly trends, daily cycles, and hourly proximity. This model mainly consists of a multi-scale decomposition layer, a parallel encoding layer, long and short distance attention modules, and an adaptive fusion layer. The multi-scale decomposition layer first models the power consumption data of silicon-based enterprises at multiple time scales and converts it into a high-dimensional feature representation. The parallel encoding layer inputs the multi-scale features into independent encoders for deep encoding and generates time-scale-specific representations. After feature processing, the adaptive fusion layer dynamically weights and fuses features at different time scales using a learnable weight matrix, outputting the final prediction result.

[0041] In one embodiment of this application, the model training employs an adaptive estimation optimizer, with a batch size of 64, an initial learning rate of 0.001, a random inactivation rate of 0.1 to prevent overfitting, a maximum number of training epochs of 300, and an early stopping strategy with a patience value of 10 epochs. Training is stopped when the validation set loss fails to improve for 10 consecutive epochs.

[0042] The training process is as follows: C1: Decompose the preprocessed power consumption time series data of silicon-based enterprises into weekly trend components, daily periodic components, and hourly proximity components. C2: Train the weekly trend component encoder, daily periodic component encoder, and hourly proximity component encoder based on the weekly trend component, daily periodic component, and hourly proximity component. C3: Based on weekly trend component encoders, daily periodic component encoders, and hourly proximity component encoders, the long-term decay characteristics, regularity characteristics, and short-term fluctuation characteristics of power consumption time series data of silicon-based enterprises are obtained. C4: Adaptive weighted fusion of the long-term decay characteristics, regularity characteristics and short-term fluctuation characteristics of the power consumption time series data of silicon-based enterprises is performed using a learnable weight matrix to obtain power consumption prediction results; C5: Based on the power consumption prediction results, the parameters of the power consumption prediction model are optimized using the mean square error loss function.

[0043] S4: Make predictions based on the trained power consumption prediction model to obtain power consumption prediction results.

[0044] In one embodiment of this application, the prediction step size of the model can be flexibly configured. For hourly datasets, the prediction step size is 1, which corresponds to predicting the power consumption of silicon-based enterprises in the next hour. For daily datasets, the prediction step size is 24, which corresponds to predicting the power consumption of silicon-based enterprises in the next day. For monthly datasets, the prediction step size is 1, which corresponds to predicting the power consumption of silicon-based enterprises in the next month. The prediction results are generated entirely based on historical data using an open-loop prediction strategy, without using previous prediction values ​​as input for subsequent time steps.

[0045] Example 2: This application provides an energy consumption prediction system for carbon footprint verification of silicon-based enterprises, including: a data acquisition module, a trend component extraction module, a periodic component extraction module and a proximity component extraction module set in parallel, a trend encoder, a periodic encoder and a proximity encoder set in parallel, and an adaptive weighted fusion module.

[0046] The data acquisition module acquires and preprocesses time-series power consumption data of silicon-based enterprises.

[0047] The trend component extraction module is used to model the long-term evolution trend of power consumption in silicon-based enterprises. This module constructs a weekly trend component from the time series data of power consumption in silicon-based enterprises, with a time window length set to 168 hours, corresponding to 7 days multiplied by 24 hours. Long-term features in the power consumption of silicon-based enterprises often contain multiple complex patterns. To effectively capture these patterns, an end-to-end trend modeling method is designed. Specifically, firstly, the original power series is linearly transformed through input value embedding; then, time embedding is introduced to model periodic patterns such as weeks and months; finally, positional encoding is used to inject the positional relationships of time steps. By summing these three embeddings element-wise, a trend representation integrating content, periodicity, and positional information can be obtained.

[0048] The periodic component extraction module is used to extract the intraday regularity and seasonality features of power consumption data from silicon-based enterprises. This module constructs daily periodic components from the time series data of power consumption of silicon-based enterprises, with a time window length set to 24 hours. The power consumption data of silicon-based enterprises simultaneously exhibits intraday periodicity and seasonal correlation. Regular changes exist between different time periods, and this periodic structure information is crucial for accurately predicting the power consumption of silicon-based enterprises. To extract effective periodic information, the original input for each day is first feature-encoded, and a multi-head attention mechanism is used to process the periodic input to extract intraday variation features. To enhance the global dependency between different dates, a positional encoding mechanism is further introduced. Finally, the periodic features are unified to the dimensions required by the model through an encoder, obtaining the periodic feature representation of each time scale at each time step.

[0049] The proximity component extraction module is used to capture the short-term, rapid fluctuations in the power consumption of silicon-based enterprises. This module constructs hourly proximity components from the time-series power consumption data of silicon-based enterprises, with time window lengths set to 1 to 6 hours. By constructing hourly time windows, rapidly changing dynamic characteristics can be captured, providing crucial information support for real-time power prediction.

[0050] The trend encoder, cycle encoder, and proximity encoder perform deep encoding of long-term evolution characteristics, regularity characteristics, and short-term fluctuation characteristics. Then, they model the dependencies between time scales to obtain weekly trend components, daily cycle components, and hourly proximity components.

[0051] The adaptive weighted fusion module includes a learnable weight module, a feature fusion module, and a prediction output module. The learnable weight module dynamically weights the three time-scale components, which are then fused by the feature fusion module. Finally, the prediction output module outputs the power consumption prediction result.

[0052] In one embodiment of this application, a learnable weight module is used to automatically learn the contribution of features at different time scales to the final prediction. This module dynamically adjusts the weight allocation through end-to-end training. A feature fusion module implements a weighted combination of the outputs of the trend encoder, period encoder, and proximity encoder, improving cross-scale feature alignment capabilities. This process enables the learning model to effectively capture important patterns and information from features at multiple time scales. The prediction output module outputs the final power consumption result for silicon-based enterprises, which comprehensively considers the combined effects of long-term trends, medium-term cycles, and short-term fluctuations. Typically, the output is represented as a prediction sequence over multiple time steps, with the prediction step size flexibly configurable according to actual needs. This framework utilizes the physical and statistical relationships between different time scales of power consumption data from silicon-based enterprises, effectively enabling joint modeling of feature representations and improving prediction accuracy and generalization performance.

[0053] Example 3: Xinjiang, as a key cluster of the silicon-based industry, possesses unique geographical advantages, with silicon resources widely distributed in areas such as Ili, Shihezi, Turpan, Hami, and Changji. During the production of industrial silicon, the rapid expansion of power consumption management systems in silicon-based enterprises has not only brought unprecedented green energy output but also placed higher demands on their power dispatch management and stable operation capabilities. Calculating energy consumption or power based on power data can comprehensively and accurately reflect macroeconomic conditions and industrial development. For example, the State Grid Corporation of China utilizes big data on electricity, combined with advanced metering and monitoring technologies, to achieve real-time monitoring and accounting of electricity consumption. This not only improves the accuracy and efficiency of accounting but also provides valuable data support for carbon footprint accounting in the silicon-based industry. Furthermore, silicon-based enterprises should provide accurate electricity consumption data to support precise calculation of power consumption. Through technological innovation and industrial upgrading, the silicon-based industry can improve energy efficiency and achieve sustainable development. Therefore, this method is validated using Xinjiang as an example.

[0054] like Figure 3 , 4 As shown in Figure 5, comparative experiments were conducted on three datasets—Yamansu, Beitashan, and Shichengzi—to verify the effectiveness of the proposed method. The results demonstrate that the proposed power consumption prediction model, MCRssFormer, achieves significant performance improvements across prediction tasks at various time scales. Figure 3 The results demonstrate the performance of MCRssFormer and iTransformer on the Yamansu dataset for a monthly silicon-based enterprise power consumption task. The black line represents the actual power value, the red line represents the prediction result of MCRssFormer, and the blue line represents the prediction result of iTransformer. It can be clearly seen that the red prediction curve of MCRssFormer is closer to the black actual value curve, especially showing stronger tracking ability in the power peak and rapidly changing regions. Figure 4 The paper presents the performance of the iTransformer and the proposed MCRssFormer model on the hourly silicon-based enterprise power consumption task on the Beitashan dataset. The black line represents the actual power value, the red line represents the prediction result of MCRssFormer, and the blue line represents the prediction result of iTransformer. It can be seen that MCRssFormer is significantly better than iTransformer in capturing short-term fluctuations and has a smaller response latency. Figure 5 The prediction performance on the Shicheng subset dataset is demonstrated, further validating the superiority and generalization ability of the method of this invention.

[0055] In one embodiment of this application, the key parameters of the model are set as follows: batch size is set to 64, initial learning rate is set to 0.001, model dimension is set to 128, number of multi-head attention heads is set to 8, number of encoder layers is set to 2, number of decoder layers is set to 1, random inactivation rate is set to 0.1, maximum number of training epochs is set to 300, and early stop strategy patience value is set to 10 epochs. In the data preprocessing stage, outlier detection and correction are first performed on the original silicon-based enterprise power consumption data. Median filtering is used to smooth the remaining noise. Then, minimum-maximum normalization is used to scale the values ​​to the range of 0 to 1. Time metadata such as holidays and days of the week is encoded using one-hot encoding. Model training uses an adaptive moment estimation optimizer, and end-to-end optimization is performed using the mean squared error loss function.

[0056] This application provides a method and system for predicting electricity consumption for carbon footprint verification of silicon-based enterprises. By constructing three parallel processing paths to capture weekly trends, daily cycles, and hourly proximity features, it effectively solves the technical bottleneck of traditional single-scale models that cannot simultaneously take into account long-term trends and short-term fluctuations. It also effectively addresses the problems of insufficient multi-scale feature capture and over-smoothing in traditional models, significantly improving the model's ability to represent the multi-scale temporal characteristics of power consumption data from silicon-based enterprises. Furthermore, the adaptive weighted fusion strategy automatically adjusts the contribution weights of features at different time scales through end-to-end learning, avoiding the subjectivity and limitations of traditional fixed-weight fusion methods. This achieves dynamic collaborative optimization of cross-scale information, enhancing the model's generalization ability and adaptability in different prediction scenarios. In addition, the encoder innovatively combines local window attention and global sparse attention through a long-short distance attention fusion mechanism. This ensures both sensitive response to local mutations and effective capture of global evolution patterns, overcoming the inherent defects of high computational complexity and over-smoothing in standard Transformer global attention, and significantly improving prediction accuracy.

[0057] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.

Claims

1. A power consumption prediction method for silicon-based enterprise carbon footprint verification, characterized in that, The method comprises the following steps: S1: obtaining and preprocessing silicon-based enterprise electricity consumption power time series data; S2: constructing an electricity consumption prediction model for silicon-based enterprise carbon footprint verification; the electricity consumption prediction model comprises a multi-scale decomposition layer, a multi-level parallel encoding and decoding layer, and an adaptive weighted fusion layer; The multi-scale decomposition layer comprises a trend component module, a periodic component extraction module, and a proximity component extraction module; the trend component module captures the long-term evolution characteristics of the silicon-based enterprise electricity consumption power time series data; the periodic component extraction module learns the regularity characteristics of the silicon-based enterprise electricity consumption power time series data; and the proximity component extraction module captures the short-term fluctuation characteristics of the silicon-based enterprise electricity consumption power time series data; The multi-level parallel encoding and decoding layer comprises a trend encoder, a periodic encoder, and a proximity encoder; the trend encoder encodes the long-term evolution characteristics in depth to obtain a weekly trend component; the periodic encoder encodes the regularity characteristics in depth to obtain a daily periodic component; and the proximity encoder encodes the short-term fluctuation characteristics in depth to obtain an hourly proximity component; The adaptive weighted fusion layer dynamically weights and fuses the three time scale components through a learnable weight matrix, and automatically adjusts the weight distribution according to the actual contribution of each scale feature to the prediction task; S3: training the electricity consumption prediction model based on the preprocessed silicon-based enterprise electricity consumption power time series data; S4: performing prediction based on the trained electricity consumption prediction model to obtain an electricity consumption prediction result.

2. The electricity consumption prediction method for silicon-based enterprise carbon footprint verification according to claim 1, wherein, The preprocessing comprises: performing preliminary screening on the original silicon-based enterprise electricity consumption power time series data by batch deletion and correction of error records, smoothing the remaining abnormal values by median filtering technology, replacing each data point with the median value of itself and adjacent observation values, scaling the numerical values by minimum-maximum normalization, and performing one-hot encoding on holiday and week time metadata.

3. The electricity consumption prediction method for silicon-based enterprise carbon footprint verification according to claim 1, wherein, The trend encoder, the periodic encoder, and the proximity encoder each comprise a long-short distance attention fusion module; The long-short distance attention fusion module comprises a short-distance attention submodule, a long-distance attention submodule, and a multi-layer perceptron; The short-distance attention submodule uniformly divides the silicon-based enterprise electricity consumption power time series data into a plurality of non-overlapping windows of fixed length, independently calculates a query matrix, a key matrix, and a value matrix within the window, models the local time dependence within the window through a scaled dot-product attention mechanism, and outputs the splicing result of the local time dependence of all windows; The long-distance attention submodule sparsely samples the silicon-based enterprise electricity consumption power time series data at a fixed step and constructs a sparse marker set, and performs global self-attention calculation on the sparse marker set; the sparse attention output is restored to the original sequence length through linear interpolation to obtain the long-distance time dependence; The multi-layer perceptron performs linear transformation and feature enhancement on the output of the short-distance attention submodule and the long-distance attention submodule.

4. The electricity consumption prediction method for silicon-based enterprise carbon footprint verification according to claim 1, wherein, The trend component module captures the long-term evolution characteristics of the silicon-based enterprise electricity consumption power time series data, including: A1: linearly transforming the silicon-based enterprise power consumption time series data of each week by input value embedding to obtain the content features of the silicon-based enterprise power consumption time series data; A2: introducing time embedding to establish periodic patterns for the silicon-based enterprise power consumption time series data of each week to obtain the periodicity of the silicon-based enterprise power consumption time series data; A3: injecting the position relationship of the time step through position encoding to obtain the position features of the silicon-based enterprise power consumption time series data; A4: fusing the content features, periodicity and position features of the silicon-based enterprise power consumption time series data to obtain the long-term evolution features of the silicon-based enterprise power consumption time series data.

5. The electricity consumption prediction method for silicon-based enterprise carbon footprint verification according to claim 1, wherein, The periodic component extraction module learns the regularity features of the silicon-based enterprise power consumption time series data, including: B1: encoding the daily silicon-based enterprise power consumption time series data; B2: using a multi-head attention mechanism to extract the intra-day variation features from the encoded data; B3: introducing a position encoding mechanism and combining the intra-day variation features to obtain the regularity features of the silicon-based enterprise power consumption time series data.

6. The electricity consumption prediction method for silicon-based enterprise carbon footprint verification according to claim 1, wherein, The window length of the short-range attention submodule and the sampling step of the long-range attention submodule include: the window length is equal to the length of the sparse sequence divided by the sampling step.

7. The electricity consumption prediction method for silicon-based enterprise carbon footprint verification according to claim 1, wherein, The adaptive weighted fusion layer further includes a loss optimization module, which uses a mean square error loss function for end-to-end training to optimize the model parameters.

8. The electricity consumption prediction method for silicon-based enterprise carbon footprint verification according to claim 1, wherein, Based on the preprocessed silicon-based enterprise power consumption time series data, the power consumption prediction model is trained, including: C1: dividing the preprocessed silicon-based enterprise power consumption time series data into week-level trend components, day-level periodic components and hour-level proximity components; C2: training trend encoders, periodic encoders and proximity encoders based on the week-level trend components, day-level periodic components and hour-level proximity components; C3: based on the trained trend encoders, periodic encoders and proximity encoders, capturing the long-term evolution features, regularity features and short-term fluctuation features of the silicon-based enterprise power consumption time series data to obtain the week-level trend components, day-level periodic components and hour-level proximity components; C4: using a learnable weight matrix to adaptively weight and fuse the week-level trend components, day-level periodic components and hour-level proximity components of the silicon-based enterprise power consumption time series data to obtain the power consumption prediction results; C5: based on the power consumption prediction results, using a mean square error loss function to optimize the parameters of the power consumption prediction model.

9. A system for electricity consumption prediction for silicon-based enterprise carbon footprint verification according to any one of claims 1-8, characterized in that, It includes: a data acquisition module, a parallelly arranged trend component extraction module, a periodic component extraction module and a proximity component extraction module, a parallelly arranged trend encoder, a periodic encoder and a proximity encoder, and an adaptive weighted fusion module; The data acquisition module acquires and preprocesses the silicon-based enterprise power consumption time series data; The trend component module captures the long-term evolution features of the silicon-based enterprise power consumption time series data; the periodic component extraction module learns the regularity features of the silicon-based enterprise power consumption time series data; the proximity component extraction module captures the short-term fluctuation features of the silicon-based enterprise power consumption time series data; The trend encoder deeply encodes long-term evolution features to obtain a weekly trend component; the periodicity encoder deeply encodes regularity features to obtain a daily periodicity component; and the proximity encoder deeply encodes short-term fluctuation features to obtain an hourly proximity component; The adaptive weighted fusion module dynamically weights and fuses the three time-scale components through a learnable weight matrix to obtain the power consumption prediction result.

10. The electricity consumption prediction system for silicon-based enterprise carbon footprint verification of claim 9, wherein, Further comprising: A visual analysis module configured to display a comparison curve of the prediction result and the true value.

Images