A prefabricated building prefabricated component production energy consumption prediction method

Abstract

This invention discloses a method for predicting energy consumption in the production of prefabricated building components, comprising the following steps: collecting and processing historical time-series data of prefabricated component production, establishing a time-series input matrix with multi-period feature enhancement; constructing a deep time-series quantile attention network, extracting local time-series features and global periodic dependencies in parallel through time-series quantile attention blocks to form a fused time-series representation; inputting the fused time-series representation into a progressive quantile decoder based on a quantile regression mechanism, outputting a component production energy consumption prediction interval with multiple confidence levels; adaptively adjusting the network configuration and training strategy according to data frequency and sample size to achieve model optimization and deployment; acquiring the target prefabricated component production time-series data to be predicted, and outputting the component production energy consumption prediction interval for the corresponding target period. This invention achieves joint optimization of prediction intervals and point predictions, and can output reliable intervals that meet the target confidence level, providing a quantitative basis for production scheduling risk decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology in industrialized building, specifically to a method for predicting energy consumption in the production of prefabricated building components. Background Technology

[0002] Currently, prefabricated construction in my country is still in its early stages. Related production, construction, and management technologies are not yet mature, and the energy-saving and carbon-reducing advantages of prefabricated buildings cannot be fully realized. Precast component production is a crucial difference between prefabricated buildings and traditional cast-in-place buildings. Accurate prediction and control of energy consumption during component production can not only provide scientific guidance for energy efficiency optimization and production scheduling in prefabricated buildings, but also fully leverage their energy-saving and carbon-reducing advantages. This is of great significance for promoting the comprehensive green transformation of my country's construction industry, facilitating energy conservation and emission reduction in the construction sector, and helping the construction industry achieve its carbon reduction goals.

[0003] Domestic and international methods for predicting industrial production energy consumption can be mainly divided into physical model methods, machine learning methods, and deep learning methods. These methods do not adequately consider the uncertainty of prediction results. The energy consumption of prefabricated component production is influenced by many factors and has significant uncertainty. It is necessary to provide the confidence level of the prediction interval so that decision-makers can obtain more comprehensive and robust decision-making information. These methods typically employ fixed network structures and training strategies. In practical applications, the sample size varies greatly among different prefabricated component manufacturers, making it difficult for the methods to adaptively adjust and limiting their generalization ability. These methods also lack sufficient fusion of local and global temporal features. For example, while the global self-attention mechanism of the standard Transformer method can capture long-distance dependencies, local changes are easily overwhelmed by global information. Recurrent neural networks, while adept at local modeling, are weak in extracting cyclical patterns across weeks and months, lacking an efficient local-global adaptive fusion mechanism. Summary of the Invention

[0004] The present invention proposes a method for predicting energy consumption in the production of prefabricated building components, which can at least solve one of the technical problems in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting energy consumption in the production of prefabricated building components includes the following steps: S1. Collect and process historical time-series data of prefabricated component production, and establish a time-series input matrix with multi-period feature enhancement; S2. Construct a deep temporal segmented attention network, extract local temporal features and global periodic dependencies in parallel through temporal segmented attention blocks, and fuse them through an adaptive gating mechanism to form a fused temporal representation; S3. Based on the quantile regression mechanism, the fused time series representation is input into the progressive quantile decoder, and the component production energy consumption prediction interval with multiple confidence levels is output. S4. Adaptively adjust network configuration and training strategy based on data frequency and sample size to achieve model optimization and deployment; S5. Obtain the production time series data of the target precast component to be predicted, input it into the optimized deployment of the deep temporal quantile attention network, and output the prediction range of component production energy consumption for the target period.

[0006] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, step S1 specifically includes: S11. Collect historical time-series data on precast component production, including component output, ambient temperature, and timestamp information; component output The output on day t, and the highest temperature. The highest and lowest ambient temperatures on day t are... The lowest ambient temperature on day t, timestamp. This is the date identifier for day t, using a standard date format, including year, month, and day fields; S12. Perform multi-period feature enhancement on the collected time-series data to construct a time-series input matrix containing original features, derived features, and periodic codes; S13. Standardize and correct outliers in the time-series input matrix.

[0007] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, step S12 specifically comprises: S121. Extracting Original Features: Arrange the year, month, and day after parsing the component production, ambient temperature, and timestamp information according to time steps to form a list of original features. , For time steps, The original number of features; S122. Calculate derived features: Calculate output volatility :

[0008] in, Let be the output volatility on day t. Let t be the component production on day t. The component production on day t-1. This is the historical average yield. Avoid division by zero; =0; Calculate the temperature-production interaction term :

[0009] in, Let t be the temperature-production interaction term on day t. This represents the highest ambient temperature on day t. The lowest ambient temperature on day t. Let t be the component output on day t; Derived feature column , For time steps, The original number of features; S123, Constructing periodic coding: Extract weekday information based on timestamp information and construct a weekday sine code. and week cosine encoding ; Identify holidays based on timestamp information and construct holiday markers. ; Periodic Encoding Column , For time steps, This is the number of periodic codes.

[0010] S124. Matrix Assembly: Align the original features, derived features, and periodic codes according to time steps, and concatenate them to form a time-series input matrix. ;

[0011] in, For the time series input matrix, For the original feature columns, For derived feature columns, This is a periodic coded column.

[0012] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, step S2 specifically includes: S21. The temporal input matrix is ​​mapped to the latent space through the embedding layer, and the position encoding of the periodic perception is superimposed to form the initial temporal representation. S22. Input the initial temporal representation into the multi-layer temporal segmentation attention block and extract hierarchical temporal features layer by layer. S23. The outputs of the last-layer temporal attention blocks are fused to form a fused temporal representation.

[0013] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, step S21 specifically comprises: S211. The feature dimensions of the temporal input matrix are mapped to the d-dimensional latent space through the embedding layer to obtain the feature embedding vector; S212. Calculate the multi-frequency sinusoidal position code and weight the frequency components at each level using learnable modulation parameters; S213. Add the feature embedding vector to the weighted position code element by element to form the initial temporal representation.

[0014] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production according to the present invention, the multi-frequency sinusoidal position coding in step S212 is constructed as follows: The required periodic scales include four periods: daily, weekly, monthly, and quarterly, with corresponding period lengths of 1 day, 7 days, 30 days, and 90 days, respectively; the latent space dimension d is evenly distributed across the four periodic scales. For the k-th period, the formula for calculating its angular frequency is:

[0015] In the formula, The period length corresponding to the k-th periodic scale. =1, =7, =30, =90; For time step and hidden space dimension The calculation formula for multi-frequency sinusoidal position coding is as follows:

[0016]

[0017] In the formula, For the position encoding matrix, For date indexing, For frequency component index, This indicates rounding down. For the position encoding matrix The Line 1 Column elements, For the position encoding matrix The Line 1 Column elements, Let d be the angular frequency of the k-th periodic scale, and d be the dimension of the latent space. The mechanism for constructing and updating learnable modulation parameters described in step S212 is as follows: Learnable modulation parameters The initialization uses normally distributed random values ​​with a mean of 1.0 and a standard deviation of 0.1, and is jointly optimized with other network parameters through backpropagation. The weighted position encoding satisfies:

[0018] In the formula, For time step Modulated position-coded vector, This indicates element-wise multiplication. For time step Modulated position-coded vector, These are learnable modulation parameters; The initial timing representation described in step S213 is calculated using the following formula:

[0019] In the formula, For time step The feature embedding vector, This is the modulated position encoding vector. This represents the initial timing characterization at time step t.

[0020] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, the time-series attention block for each layer in step S22 includes the following parallel processing branches: S221. Parallel computation of dilated causal convolution branches, expanding the receptive field through adaptive dilation factor to extract local temporal features; S222: Parallel computation of multi-head self-attention branches, extracting global periodic dependencies through causal mask constraints; S223. Adaptively fuse local temporal features and global periodic dependencies through a gating mechanism to generate fused temporal features; S224. Connect the fused temporal features with the input residuals, and output the result after layer normalization. Specifically, the gating mechanism described in step S223 is as follows: Calculate the gating weights:

[0021] in, For date Local temporal characteristics, For date The global periodic dependency feature, For bias vectors, ; Generate fusion features:

[0022] in For date Gating weights, For date Local temporal characteristics, For date The global periodic dependency feature, This indicates element-wise multiplication.

[0023] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, step S3 specifically comprises: S31. Input the fused temporal representation into the progressive quantile decoder, extract the hidden state of the last time step as the decoding starting point, and generate the initial quantile prediction. S32. Iteratively optimize quantile prediction through residual refinement mechanism; S33. Apply monotonicity constraints to ensure that the lower quantile, median prediction, and upper quantile increase sequentially, and output the component production energy consumption prediction range.

[0024] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, the residual refinement mechanism in step S32 has the following iterative formula:

[0025] Where k is the iteration step, For date The residual at the k-th iteration step For date The residual at the (k-1)th iteration step For the first k The refinement of the network step by step, This is the hidden state of the last time step output of the final time-sequence attention block.

[0026] As a preferred embodiment of the energy consumption prediction method for prefabricated building component production described in this invention, step S4 specifically comprises: S41. Adaptively select network configuration based on data time resolution: use standard attention and moderate network depth for daily data, and use sparse attention and deeper network depth for hourly data. S42. Adaptively select training strategies based on sample size: use strong regularization and short sequence input for scenarios with few samples, and use deep networks and long sequence input for scenarios with many samples. S43. Based on validation set performance monitoring, trigger the learning rate adjustment and early stopping mechanism to complete model training.

[0027] The beneficial effects of this invention are: Compared with traditional methods that construct intervals after the fact or use fixed widths, this invention achieves joint optimization of prediction intervals and point predictions, and can output reliable intervals that meet the target confidence level, providing a quantitative basis for production scheduling risk decisions. This invention adaptively selects the attention mechanism type (standard / sparse), network depth (number of layers), and regularization strength based on the data sample size, and automatically triggers learning rate adjustment and early stopping based on validation set performance. This adaptive mechanism enables the same network framework to adapt to data conditions of different component factories, overcoming the shortcomings of existing models that suffer from significant performance fluctuations due to differences in data sample size, and exhibits good engineering generalization ability. This invention designs a temporal quantized attention block that computes dilated causal convolution (extracting local recent changes) and multi-head logarithmic sparse self-attention (capturing long-range dependencies across weeks and months) in parallel, and dynamically fuses the outputs of the two through a learnable gating mechanism, thereby achieving adaptive fusion of local and global temporal features. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the method steps of the energy consumption prediction method for prefabricated building component production according to the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0030] like Figure 1 As shown, a method for predicting energy consumption in the production of prefabricated building components includes the following steps: S1. Collect and process historical time-series data of prefabricated component production, and establish a time-series input matrix with multi-period feature enhancement; S2. Construct a deep temporal segmented attention network, extract local temporal features and global periodic dependencies in parallel through temporal segmented attention blocks, and fuse them through an adaptive gating mechanism to form a fused temporal representation; S3. Based on the quantile regression mechanism, the fused time series representation is input into the progressive quantile decoder, and the output is the component production energy consumption prediction range with multiple confidence levels. S4. Adaptively adjust network configuration and training strategy based on data frequency and sample size to achieve model optimization and deployment; S5. Obtain the production time series data of the target precast component to be predicted, input it into the optimized deployment of the deep temporal quantile attention network, and output the prediction range of component production energy consumption for the target period.

[0031] Step S1 is as follows: S11. Collect historical time-series data on precast component production, including component output, ambient temperature, and timestamp information. Component output... Specifically, the volume (m³) of prefabricated components produced on day t, and the ambient temperature is specifically the highest temperature of that day. and lowest temperature timestamp information After parsing, the year, month, and day fields are extracted, and the weekday number, month number, quarter number, and cumulative number of days in the year are calculated. Specifically, for example, the parsing result of the timestamp information 2024-03-15 is the year. ,month ,date Weekday number (Friday), Quarterly Sequence (First Quarter), Annual Cumulative Days (The 75th day of 2024, a leap year) (Assuming it is not a holiday); S12. Perform multi-period feature enhancement on the collected time-series data to construct a time-series input matrix containing original features, derived features, and periodic codes; S13. Standardize and correct outliers in the time series input matrix; Step S12 is as follows: S121. Extracting Original Features: Arrange the year, month, and day after parsing the component production, ambient temperature, and timestamp information according to time steps to form a list of original features. , For time steps, These are the original feature numbers.

[0032]

[0033] S122. Calculate derived features: Calculate output volatility :

[0034] in, Let be the output volatility on day t. Let t be the component production on day t. The component production on day t-1. This is the historical average yield. Avoid division by zero; =0; Calculate the temperature-production interaction term :

[0035] in, Let t be the temperature-production interaction term on day t. This represents the highest ambient temperature on day t. The lowest ambient temperature on day t. Let t be the component output on day t; Derived feature column , For time steps, The original number of features;

[0036] S123, Constructing periodic coding: Extract weekday information based on timestamp information and construct a weekday sine code. and week cosine encoding :

[0037] Identify holidays based on timestamp information and construct holiday markers. :

[0038] Periodic Encoding Column , For time steps, For the number of periodic codes;

[0039] S124. Matrix Assembly: Align the original features, derived features, and periodic codes according to time steps, and concatenate them to form a time-series input matrix. ;

[0040] Step S13 is as follows: Z-score standardization:

[0041] In the formula, For the first t day j The original value of the feature. For the first j The mean of the features, For the first j Standard deviation of features To avoid dividing by zero for small constants; 5 Criteria for identifying outliers:

[0042] in, and The first jThe mean and standard deviation of each feature column; Lagrange interpolation repair: For those judged as abnormal or missing Select k nearest normal data points (usually k=3 or 5) and construct an interpolation polynomial:

[0043]

[0044] The time-series input matrix is ​​standardized and cleaned.

[0045]

[0046] in, Avoid division by zero; Step S2 is as follows: S21. The temporal input matrix is ​​mapped to the latent space through the embedding layer, and the position encoding of the periodic perception is superimposed to form the initial temporal representation. S22. Input the initial temporal representation into the multi-layer temporal segmentation attention block and extract hierarchical temporal features layer by layer. S23. The outputs of the last-layer temporal attention blocks are fused to form a fused temporal representation.

[0047] Step S21 is as follows: S211, The timing input matrix is ​​transmitted through the embedding layer. Mapped to d The latent space is used to obtain the feature embedding vectors at each time step. t Feature embedding vector

[0048]

[0049] in: For learnable embedding matrices, For bias vectors, d =64 represents the implicit space dimension.

[0050] S212. Calculate the multi-frequency sinusoidal position code and weight the frequency components at each level using learnable modulation parameters; The latent space dimension d is evenly distributed across four periodic scales: daily, weekly, monthly, and quarterly. For example, when d=128, the periodic scale... The corresponding latent space dimension intervals are [0,31], [32,63], [64,95], and [96,127], respectively. The angular frequency of the k-th period scale is:

[0051] In the formula, Let be the angular frequency of the k-th period scale. The period length corresponding to the k-th periodic scale. =1, =7, =30, =90.

[0052] The period length corresponding to the k-th periodic scale. =1, =7, =30, =90.

[0053] Multi-frequency sinusoidal position coding matrix The calculation is as follows:

[0054]

[0055] In the formula, For the position encoding matrix, For date indexing, For frequency component index, This indicates rounding down. For the position encoding matrix The Line 1 Column elements, For the position encoding matrix The Line 1 Column elements, Let d be the angular frequency of the k-th periodic scale, and d be the dimension of the latent space.

[0056] Traditional Transformer position encoding uses fixed parameters, which cannot adapt to the actual periodic characteristics of data. This embodiment introduces learnable modulation parameters. Adaptive weighting of each frequency component:

[0057] In the formula, For time step Modulated position-coded vector, This indicates element-wise multiplication. For time step Modulated position-coded vector, These are learnable modulation parameters.

[0058] The learning mechanism for modulation parameters is as follows: First, during the initialization phase, let... Secondly, the forward propagation stage calculates the weighted positional encoding and superimposes it with the feature embedding; thirdly, the backpropagation stage predicts the error gradient. The parameters are updated; finally, during the convergence phase, the frequency components that match the actual data cycle obtain larger absolute values, and irrelevant frequencies are suppressed, achieving data-driven cycle awareness. For example, if the training data exhibits strong periodicity (such as a sudden drop in energy consumption due to weekend production stoppages), the modulation parameter of the corresponding weekly frequency automatically increases; if a certain cycle scale is irrelevant to the actual data (such as no significant quarterly pattern), the modulation parameter of the corresponding dimension will decay to near zero, achieving automatic suppression.

[0059] S213. Add the feature embedding vector to the weighted position code element by element to form the initial temporal representation;

[0060] Synergistic explanation with step S213: The artificially constructed periodic encoding in step S123 serves as an explicit prior input network; the learnable periodic position encoding in step S21 serves as an implicit enhancement mechanism. Together, they form an explicit-implicit dual-channel periodic modeling, improving prediction robustness.

[0061] Step S22 inputs the initial temporal representation into a multi-layer (e.g., L=4 layers) temporal quantile attention block, and extracts hierarchical temporal features layer by layer; each layer of temporal quantile attention block contains the following parallel processing branches: S221. Parallel computation of dilated causal convolution branches, expanding the receptive field through adaptive dilation factor to extract local temporal features; For the Layered temporal attention block, input is ,in For the first t The input hidden states at each time step; the dilated causal convolution operation is defined as:

[0062] in, The kernel size is [size]. As a void factor, For the first Each convolutional kernel weight matrix For the bias vector; ReLU( () is an element-wise activation function, applied to each component of the input vector. calculate ; Local temporal features are ensured through causal filling It relies solely on information from the current and historical time steps.

[0063] The adaptive dilation factor is configured as [2, 4, 8, 16], the convolutional kernel size is set to 3, and the receptive field covers at least 30 days to capture monthly cyclical patterns; the receptive field is expanded by the adaptive dilation factor to extract local temporal features. .

[0064] S222: Parallel computation of multi-head self-attention branches, extracting global periodic dependencies through causal mask constraints; use The attention heads are computed in parallel, and the dimension of each head is... .

[0065] For the Size: , ,

[0066] in, , , It is a learnable projection matrix.

[0067] Attention weight calculation introduces causal masking The definition is as follows:

[0068]

[0069]

[0070] Will The outputs of each size are spliced ​​and linearly projected:

[0071] in, To output the projection matrix; To further reduce computational overhead, a LogSparse sparsity strategy is adopted, focusing only on historical positions with a logarithmic interval for each location, thus reducing computational complexity from... Down to .

[0072] S223. Adaptively fuse local temporal features and global periodic dependencies through a gating mechanism to generate fused temporal features.

[0073] Local temporal features Global periodicity dependency features The data is then concatenated, and adaptive weights are calculated using a gating network.

[0074] in, For the gated weight matrix, For bias vectors, ; Generate fused temporal features:

[0075] Here, ⊙ represents element-wise multiplication. A value approaching 1 indicates that local factors dominate, while a value approaching 0 indicates that the global cycle dominates.

[0076] S224. The fused temporal features are concatenated with the input residuals and output after layer normalization.

[0077] Concatenate the fused temporal features with the input residuals:

[0078] Perform layer normalization:

[0079] in, and These are the mean and standard deviation of the input to this layer along the feature dimension, respectively. , For learnable scaling and translation parameters, Here is the numerical stability constant. This is a layer index.

[0080] S23. The outputs of the last-layer temporal attention blocks are fused to form a fused temporal representation.

[0081] The output of the temporal segmentation attention block in the last layer (the Lth layer) is used as the fused temporal representation. This is used for subsequent quantile decoding.

[0082] Step S3 is as follows: S31. Input the fused temporal representation into the progressive quantization decoder to extract the hidden state of the last time step of the fused temporal representation. As the starting point for decoding; median predictions are generated through a basic prediction network (e.g., a two-layer fully connected network). Non-negative offsets are generated through an offset prediction network (with ReLU activation to ensure non-negativity). Generate initial quantile predictions.

[0083]

[0084] in, The lower quantile level (e.g., 0.1). The upper quantile level (e.g., 0.9).

[0085] S32. Iteratively optimize quantile prediction using a residual refinement mechanism. The formula for the k-th iteration is:

[0086] In the formula, k For the iteration step, This is the hidden state of the last time step output of the final time-sequence attention block. For the first k The refined network of the next step takes as input a concatenated vector of the final hidden state and the prediction from the previous iteration. The number of iterations is... k It can be set to 3 to 5 steps.

[0087] S33. Apply monotonicity constraints to ensure that the lower quantile, median prediction, and upper quantile increase sequentially. The lower quantile is generated by subtracting the offset from the median prediction, and the upper quantile is generated by adding the offset to the median prediction. Output the component production energy consumption prediction interval, including the lower quantile. Median prediction Upper quantile These correspond to optimistic estimation, most likely estimation, and pessimistic estimation, respectively.

[0088] Step S4 is as follows: S41. Adaptively select network configuration based on data time resolution: use standard attention and moderate network depth for daily data, and use sparse attention and deeper network depth for hourly data. S42. Adaptively select training strategies based on sample size: use strong regularization and short sequence input for scenarios with few samples, and use deep networks and long sequence input for scenarios with many samples. S43. Based on validation set performance monitoring, trigger the learning rate adjustment and early stopping mechanism to complete model optimization.

[0089] Step S5 is as follows: The production time series data of the target precast component to be predicted is obtained, and then input into the trained deep temporal quantile attention network to output the predicted energy consumption range of component production for the target period.

[0090] This embodiment takes the prediction of production energy consumption of a prefabricated building component factory as an example to further illustrate the implementation process of the method for predicting production energy consumption of prefabricated building components based on deep temporal quantile attention network.

[0091] (1) Collect and process historical time series data of prefabricated component production, and establish a time series input matrix with multi-period feature enhancement; Production data from the component factory was collected, with a daily time resolution, totaling 577 samples. This data includes the original variables: component output, highest temperature, lowest temperature, and timestamp information. A time-series input matrix was constructed with a feature dimension F=10, including the original features (component output). highest temperature minimum temperature ,Year ,moon Derivative characteristics (output volatility) Temperature-related interactive items Periodic encoding (weekday sine encoding) Week cosine encoding Holiday markers ); Z-score standardization was adopted, outliers were corrected based on the 5σ criterion, and Lagrange interpolation was used for patching. (2) Construct a deep temporal segmented attention network, extract local temporal features and global periodic dependencies in parallel through temporal segmented attention blocks, and fuse them through an adaptive gating mechanism to form a fused temporal representation; The specific configuration of the deep temporal fractional attention network is as follows: The latent space dimension d=128, which is evenly distributed across four periodic scales, with 32 dimensions per scale; the number of temporal quantile attention block layers L=4; the number of multi-head self-attention heads... =8, dimensions per head =16; Diffuse causal convolution kernel size M=3; Sigmoid activation is used for the gated fusion layer; (3) Based on the quantile regression mechanism, the fused time series representation is input into the progressive quantile decoder, and the output is the component production energy consumption prediction interval with multiple confidence levels; The progressive quantization decoder is specifically configured as follows: The number of iterations is k=3; the quantile level set is Q={0.1,0.5,0.9}. (4) Based on the data frequency and sample size, adaptively adjust the network configuration and training strategy to achieve model optimization and deployment; The optimizer uses Adam with an initial learning rate of 1× Weight decay 1× ; (5) Obtain the production time series data of the target prefabricated component to be predicted, input it into the optimized deployment of the deep time series quantile attention network, and output the component production energy consumption prediction interval for the target period.

[0092] The collected production data resulted in a predicted energy consumption range for component production of {3154.13, 4506.89, 5642.5}, while the actual energy consumption was 4671.47 kW. h, the predicted energy consumption range for component production successfully covered the actual energy consumption, with a relative error of 3.65% between the predicted median and the actual energy consumption.

[0093] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0094] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting energy consumption in production of prefabricated components for fabricated buildings, characterized in that, Includes the following steps: S1. Collect and process historical time-series data of prefabricated component production, and establish a time-series input matrix with multi-period feature enhancement; S2. Construct a deep temporal segmented attention network, extract local temporal features and global periodic dependencies in parallel through temporal segmented attention blocks, and fuse them through an adaptive gating mechanism to form a fused temporal representation; S3. Based on the quantile regression mechanism, the fused time series representation is input into the progressive quantile decoder, and the component production energy consumption prediction interval with multiple confidence levels is output. S4. Adaptively adjust network configuration and training strategy based on data frequency and sample size to achieve model optimization and deployment; S5. Obtain the production time series data of the target precast component to be predicted, input it into the optimized deployment of the deep temporal quantile attention network, and output the prediction range of component production energy consumption for the target period. 2.The method according to claim 1, characterized in that: Step S1 specifically includes: S11. Collect historical time-series data on precast component production, including component output, ambient temperature, and timestamp information; component output The output on day t, and the highest temperature. The highest and lowest ambient temperatures on day t are... The lowest ambient temperature on day t, timestamp. This is the date identifier for day t, using a standard date format, including year, month, and day fields; S12. Perform multi-period feature enhancement on the collected time-series data to construct a time-series input matrix containing original features, derived features, and periodic codes; S13. Standardize and correct outliers in the time-series input matrix.

3. The method for predicting energy consumption in the production of prefabricated building components according to claim 2, characterized in that: Step S12 is as follows: S121. Extracting Original Features: Arrange the year, month, and day after parsing the component production, ambient temperature, and timestamp information according to time steps to form a list of original features. , For time steps, The original number of features; S122. Calculate derived features: Calculate output volatility : in, Let be the output volatility on day t. Let t be the component production on day t. The component production on day t-1. This is the historical average yield. Avoid division by zero; =0; Calculate the temperature-production interaction term : in, Let t be the temperature-production interaction term on day t. This represents the highest ambient temperature on day t. The lowest ambient temperature on day t. Let t be the component output on day t; Derived feature column , For time steps, The original number of features; S123, Constructing periodic coding: Extract weekday information based on timestamp information and construct a weekday sine code. and week cosine encoding ; Identify holidays based on timestamp information and construct holiday markers. ; Periodic Encoding Column , For time steps, This is the number of periodic codes. S124. Matrix Assembly: Align the original features, derived features, and periodic codes according to time steps, and concatenate them to form a time-series input matrix. ; in, For the time series input matrix, For the original feature columns, For derived feature columns, This is a periodic coded column.

4. The method for predicting energy consumption in the production of prefabricated building components according to claim 1, characterized in that: Step S2 specifically includes: S21. The temporal input matrix is ​​mapped to the latent space through the embedding layer, and the position encoding of the periodic perception is superimposed to form the initial temporal representation. S22. Input the initial temporal representation into the multi-layer temporal segmentation attention block and extract hierarchical temporal features layer by layer. S23. The outputs of the last-layer temporal attention blocks are fused to form a fused temporal representation.

5. The method for predicting energy consumption in the production of prefabricated components for prefabricated buildings according to claim 4, characterized in that: Step S21 is as follows: S211. The feature dimensions of the temporal input matrix are mapped to the d-dimensional latent space through the embedding layer to obtain the feature embedding vector; S212. Calculate the multi-frequency sinusoidal position code and weight the frequency components at each level using learnable modulation parameters; S213. Add the feature embedding vector to the weighted position code element by element to form the initial temporal representation.

6. The method for predicting energy consumption in the production of prefabricated building components according to claim 5, characterized in that: The multi-frequency sinusoidal position coding described in step S212 is constructed as follows: The required periodic scales include four periods: daily, weekly, monthly, and quarterly, with corresponding period lengths of 1 day, 7 days, 30 days, and 90 days, respectively; the latent space dimension d is evenly distributed across the four periodic scales. For the k-th period, the formula for calculating its angular frequency is: In the formula, The period length corresponding to the k-th periodic scale. =1, =7, =30, =90; For time step and hidden space dimension The calculation formula for multi-frequency sinusoidal position coding is as follows: In the formula, For the position encoding matrix, For date indexing, For frequency component index, Indicates rounding down. For the position encoding matrix The Line number Column elements, For the position encoding matrix The Line number Column elements, Let d be the angular frequency of the k-th periodic scale, and d be the dimension of the latent space. The mechanism for constructing and updating learnable modulation parameters described in step S212 is as follows: Learnable modulation parameters The initialization uses normally distributed random values ​​with a mean of 1.0 and a standard deviation of 0.1, and is jointly optimized with other network parameters through backpropagation. The weighted position encoding satisfies: In the formula, For time steps Modulated position-coded vector, This indicates element-wise multiplication. For time step Modulated position-coded vector, These are learnable modulation parameters; The initial timing representation described in step S213 is calculated using the following formula: In the formula, For time step The feature embedding vector, This is the modulated position encoding vector. This represents the initial timing characterization at time step t.

7. The method for predicting energy consumption in the production of prefabricated building components according to claim 4, characterized in that: Each temporal segmentation attention block in step S22 includes the following parallel processing branches: S221. Parallel computation of dilated causal convolution branches, expanding the receptive field through adaptive dilation factor to extract local temporal features; S222: Parallel computation of multi-head self-attention branches, extracting global periodic dependencies through causal mask constraints; S223. Adaptively fuse local temporal features and global periodic dependencies through a gating mechanism to generate fused temporal features; S224. Connect the fused temporal features with the input residuals, and output the result after layer normalization. Specifically, the gating mechanism described in step S223 is as follows: Calculate the gating weights: in, For date Local temporal characteristics, For date The global periodic dependency feature, For bias vectors, ; Generate fusion features: in For date Gating weights, For date Local temporal characteristics, For date The global periodic dependency feature, This indicates element-wise multiplication.

8. The method for predicting energy consumption in the production of prefabricated components for prefabricated buildings according to claim 1, characterized in that: Step S3 is as follows: S31. Input the fused temporal representation into the progressive quantile decoder, extract the hidden state of the last time step as the decoding starting point, and generate the initial quantile prediction. S32. Iteratively optimize quantile prediction through residual refinement mechanism; S33. Apply monotonicity constraints to ensure that the lower quantile, median prediction, and upper quantile increase sequentially, and output the component production energy consumption prediction range.

9. The method for predicting energy consumption in the production of prefabricated components for prefabricated buildings according to claim 8, characterized in that: The residual refinement mechanism described in step S32 has the following iterative formula: Where k is the iteration step, For date The residual at the k-th iteration step For date The residual at the (k-1)th iteration step For the first k The refinement of the network step by step, This is the hidden state of the last time step output of the final time-sequence attention block.

10. The method for predicting energy consumption in the production of prefabricated components for prefabricated buildings according to claim 1, characterized in that: Step S4 is as follows: S41. Adaptively select network configuration based on data time resolution: use standard attention and moderate network depth for daily data, and use sparse attention and deeper network depth for hourly data. S42. Adaptively select training strategies based on sample size: use strong regularization and short sequence input for scenarios with few samples, and use deep networks and long sequence input for scenarios with many samples. S43. Based on validation set performance monitoring, trigger the learning rate adjustment and early stopping mechanism to complete model training.

Images