Intelligent energy-saving facility operation evaluation method based on artificial intelligence

Abstract

The invention discloses an intelligent evaluation method for operation of an energy-saving facility based on artificial intelligence. The method comprises the following steps: step 1, collecting multi-source time sequence data of the energy-saving facility; 2, extracting a supervision label and binding the supervision label with the paired sample; step 3, inputting the paired samples into an improved TSMixer model, generating feature mixing weight parameters and bias parameters in a feature mixing layer through a HyperNetwork sub-module, and embedding an SE-block channel attention structure to obtain intermediate feature representation; 4, the step 3 is executed repeatedly; 5, constructing a dynamic feature fusion network; 6, inputting the comprehensive feature set into the LightGBM model, and outputting a result sequence; and 7, comparing the result sequence with the corresponding reference data in the reference library, and outputting an energy-saving facility operation intelligent evaluation result. The method has the advantages of comprehensive evaluation, efficient processing and intelligent result.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and energy-saving control technology, and in particular to an intelligent evaluation method for the operation of energy-saving facilities based on artificial intelligence. Background Technology

[0002] With the deepening of the concept of energy conservation and environmental protection, various energy-saving facilities are being used in industry, construction, transportation, and other fields.

[0003] With widespread deployment in this field, its operational efficiency and energy consumption levels have become crucial indicators affecting system performance and operating costs. Traditional energy-saving assessment methods largely rely on manually set rules or static threshold judgments, making it difficult to accurately reflect the true performance of equipment under complex operating conditions. Furthermore, when faced with multi-source heterogeneous time-series data, such as equipment operating data, environmental information, and load fluctuations, existing methods have significant limitations in data fusion and dynamic modeling, failing to achieve comprehensive modeling and intelligent assessment of equipment energy efficiency, status, and potential anomalies.

[0004] With the development of machine learning and deep learning technologies, data-driven energy consumption prediction and operational analysis methods have gradually become mainstream. Some methods introduce neural network models to model time series data or use ensemble learning algorithms for state recognition, but these still suffer from insufficient feature representation capabilities, weak contextual correlation, and limited evaluation dimensions. Especially in the context of high-dimensional, multi-period changing data, effectively integrating dynamic time information with static environmental factors to establish a universal and adaptable intelligent evaluation mechanism remains a key research challenge.

[0005] Most current assessment systems lack targeted data label definitions and monitoring mechanisms, making it difficult to simultaneously evaluate operational indicators across different dimensions, such as energy consumption levels, operational status scores, and abnormal behaviors. Furthermore, the interpretability and traceability of assessment results are poor, lacking effective comparison mechanisms with historical operational samples and industry standards. This results in assessment conclusions lacking verifiable evidence and are difficult to use for operational decision support. While some methods propose feature enhancement strategies, they remain insufficient for multi-scale feature fusion under dynamically changing operating conditions, and in particular, lack stable and efficient model structures to support multi-stage, reusable assessment processes.

[0006] Therefore, how to provide an intelligent assessment method for the operation of energy-saving facilities based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent assessment method for the operation of energy-saving facilities based on artificial intelligence. This invention fully utilizes the improved TSMixer deep model, HyperNetwork dynamic mapping mechanism, SE-block attention structure and LightGBM multi-task prediction algorithm, and describes in detail the feature extraction, intelligent prediction and benchmark comparison process of the operating status of energy-saving facilities. It has the advantages of accurate assessment, fast response, strong adaptability and high scalability.

[0008] An intelligent assessment method for the operation of energy-saving facilities based on artificial intelligence, according to an embodiment of the present invention, includes the following steps:

[0009] Step 1: Collect multi-source time series data of energy-saving facilities, preprocess the multi-source time series data, and generate context feature vectors;

[0010] Step 2: Extract time step sequences from the multi-source aligned time series, select corresponding context feature vectors for each time step sequence for pairing, and extract supervision labels from the target variable time series according to the preset prediction time range and bind them to the paired samples;

[0011] Step 3: Input the paired samples into the improved TSMixer model, map the time step sequence in the temporal mixing layer, generate feature mixing weight parameters and bias parameters through HyperNetwork with context feature vector as input and adjust the channel dimension mapping relationship in the feature mixing layer, embed the SE-block channel attention structure into the output of the feature mixing layer and combine residual connections and layer normalization to obtain intermediate feature representations;

[0012] Step 4: Repeat Step 3 according to the preset number of layers to obtain the final output features;

[0013] Step 5: Aggregate the final output features along the time dimension and construct a dynamic feature fusion network with the corresponding context feature vectors to form a comprehensive feature set;

[0014] Step 6: Input the comprehensive feature set into the LightGBM model and output the energy consumption prediction results, operating status score and anomaly indication;

[0015] Step 7: Compare the result sequence with the corresponding benchmark data in the benchmark library, and output the intelligent assessment results of energy-saving facility operation.

[0016] Optionally, step one specifically includes:

[0017] Multi-source time series data includes equipment operation data, environmental data, and load data;

[0018] Perform timestamp alignment and synchronization on multi-source time series data, and perform missing value processing, noise suppression, and outlier sample removal;

[0019] Units and dimensions are unified, numerical values ​​are normalized and categorically encoded, and time, environmental and operational status features are extracted to generate context feature vectors.

[0020] Optionally, step two specifically includes:

[0021] Set the window length, sliding step size, and preset prediction time range for the sliding window;

[0022] Extract time step sequences from multi-source aligned time series according to window length and sliding step size to obtain candidate time step sequences;

[0023] For each candidate time step sequence, a corresponding context feature vector is selected according to the timestamp association method to form paired samples;

[0024] Determine the time series of the target variable, and extract the supervision labels corresponding to each paired sample from the time series of the target variable according to the preset prediction time range, and bind the supervision labels to the paired samples;

[0025] Candidate samples that do not meet the window integrity requirement will be removed, and candidate samples that cannot obtain supervision labels due to exceeding the preset prediction time range will also be removed.

[0026] Record the start and end timestamps, target variable identifiers, and context feature vector indices for each paired sample to generate a sample set.

[0027] Optionally, step three specifically includes:

[0028] Paired samples are input into the improved TSMixer model, and the paired samples include time step sequences and corresponding context feature vectors;

[0029] The time-step sequence is mapped in the time-mixing layer and the time-mixing features are output.

[0030] In the feature fusion layer, the context feature vector is used as input to the HyperNetwork to generate feature fusion weight parameters and bias parameters;

[0031] Based on the feature mixing weight parameters, a linear mapping is performed on the feature variables corresponding to the temporal mixed features at each time step. The weighted summation is performed according to the channel dimension and the bias parameters are superimposed to generate the channel mapping result.

[0032] The channel mapping results are subjected to dimension restoration and shape recovery to obtain feature fusion output;

[0033] For each channel of the feature blending output, an averaging operation is performed in the time dimension. The feature values ​​at each time step are summed and divided by the number of time steps to obtain the average feature value of the current channel. The average feature values ​​of each channel are combined in channel order to form the global average pooling result.

[0034] An SE-block channel attention structure is embedded in the output of the feature mixing layer, and channel weights are calculated based on the global average pooling results in the time dimension.

[0035] The intermediate feature representation is obtained by weighting and recalibrating the features of each channel on the output of the feature mixing layer using channel weights, and combining residual connections and layer normalization.

[0036] Optionally, step four specifically includes: setting a preset number of layers and layer order index, using the intermediate feature representation as the layer input feature of the first layer; repeating the processing flow described in step three for each layer; using the intermediate feature representation of the current layer as the layer input feature of the next layer and continuing to execute the processing flow described in step three until all preset number of layers are completed; and using the intermediate feature representation of the last layer as the final output feature of the improved TSMixer model.

[0037] Optionally, step five specifically includes:

[0038] The final output features are aggregated in the time dimension, and the mean, extreme values ​​and rate of change within the window are calculated to form a time-aggregated representation.

[0039] A dynamic feature fusion network is constructed by taking the temporal aggregation representation and the corresponding context feature vector as input.

[0040] The dynamic feature fusion network includes a random feature aggregation module and a dimension mixing module;

[0041] In the random feature aggregation module, N sub-representations are extracted from the temporal aggregation representation based on a random sampling strategy. During the inference stage, the sub-representations are weighted and converged to generate core features.

[0042] The core features are integrated with the contextual feature vectors to form a fused input;

[0043] In the dimension blending module, a scalable dimension transformation is performed on the fusion input, and the output is a fusion feature representation;

[0044] Layer normalization and dimensional integration are performed on the fused feature representation to form a comprehensive feature set.

[0045] Optionally, the step of extracting N sub-representations from the temporal aggregate representation based on a random sampling strategy, and performing weighted convergence on the sub-representations to generate core features during the inference phase, specifically includes:

[0046] First, set the number of samples, random seed, and alignment specifications for the target length and number of channels of the time aggregation representation, and establish an index set for the time dimension and the channel dimension;

[0047] M sets of random masks are generated based on the set number of samplings. Each set of masks includes a time sub-interval index and a channel subset index.

[0048] During each sampling process, the corresponding time segment and channel subset are selected from the time aggregation representation based on a random mask;

[0049] Each time interval and channel subset together constitute a sub-representation;

[0050] Perform length pruning or zero-padding on each sub-representation to keep all sub-representations consistent in time step length, and perform mean removal and amplitude normalization on the channel dimension to align the scale.

[0051] The size-aligned sub-representations are input into the encoding unit, where sub-representation vectors are generated and stored through linear transformations and nonlinear activations.

[0052] During the inference phase, all sub-representation vectors are collected and weights are generated in a preset order to form the input. The fusion weights are calculated for each sub-representation, and the fusion weights are nonnegated and normalized.

[0053] All sub-representations are weighted and summed according to their respective weights to obtain a weighted combination result, which serves as the basic representation of the core feature.

[0054] The weighted results are then subjected to dimensionality restoration and shape recovery, mapped to the same tensor structure as the time aggregation representation, to obtain the core features.

[0055] Optionally, step six specifically includes:

[0056] Energy consumption labels, operational status rating labels, and anomaly indication labels are extracted from the sample set according to the definition of supervision labels.

[0057] The aligned sample set is divided into training set, validation set and test set according to time order and proportion;

[0058] Construct sub-models, including a LightGBM regression sub-model for energy consumption prediction, a LightGBM regression sub-model for operational status scoring, and a LightGBM classification sub-model for anomaly indication;

[0059] Set training parameters for each sub-model, including number of iterations, learning rate, maximum depth, number of leaf nodes, subsample ratio, feature sampling ratio, and regularization coefficient;

[0060] Each sub-model is trained separately on the training set, and the number of iterations and loss convergence are monitored on the validation set to determine the optimal number of iterations.

[0061] Perform inference on the test set for each sub-model corresponding to the optimal number of iterations and output the raw output of energy consumption prediction results, running status score and anomaly indication;

[0062] The original outputs of each sub-model are timestamped according to the sample index to form a result sequence that corresponds one-to-one with the comprehensive feature set;

[0063] The result sequence includes energy consumption prediction results, operating status scores, and anomaly indicators.

[0064] Optionally, step seven specifically includes:

[0065] Read the benchmark record corresponding to the sample set from the benchmark library;

[0066] The baseline records are retrieved based on the timestamps, device identifiers, and contextual feature vectors in the result sequence, and the baseline data that is consistent with the result sequence in terms of device type, load level, environmental conditions, and time segmentation is selected.

[0067] Align the resulting sequence with the baseline data on the time axis and unify the units of measurement and dimensions; calculate the difference measure and relative difference measure between the energy consumption prediction results and the corresponding baseline energy consumption values ​​to obtain the energy consumption deviation sequence.

[0068] The difference measure and relative difference measure are calculated for the operating status score and the corresponding benchmark score, respectively, to obtain the status score deviation sequence;

[0069] Based on the anomaly indications and the running status rules in the benchmark library, a time-by-time matching is performed to generate an anomaly matching marker sequence;

[0070] Within a preset time window, calculate window statistics for the energy consumption deviation sequence and the state score deviation sequence, and determine dynamic threshold parameters based on threshold rules to determine whether the deviation exceeds the limit within the window.

[0071] The energy consumption deviation sequence, the state score deviation sequence, and the abnormal matching mark sequence are weighted and integrated according to preset weights to form an evaluation judgment vector;

[0072] Based on the assessment judgment vector and threshold rules, intelligent assessment results of energy-saving facility operation are generated. The assessment results include assessment level identifier, out-of-bounds time period, associated equipment identifier and corresponding context feature vector index.

[0073] The assessment results are archived in chronological order and with equipment identification, and the intelligent assessment results of energy-saving facility operation are output.

[0074] The beneficial effects of this invention are:

[0075] This invention introduces an improved TSMixer model and a HyperNetwork dynamic adjustment mechanism, combined with the collaborative processing of time mixing layers and feature mixing layers, to achieve deep coupling modeling between multi-source time series data and contextual information, effectively improving feature representation capabilities and modeling flexibility.

[0076] This invention innovatively constructs a dynamic feature fusion network, introducing random feature aggregation and dimensionality mixing modules in the fusion stage. Core features are generated through multiple sub-representation extractions and weighted convergence, and contextual feature vectors are further fused, making the final comprehensive features more discriminative and generalizable, providing a more stable and reliable input for downstream tasks.

[0077] This invention designs a prediction structure based on a multi-task LightGBM sub-model to achieve joint modeling and output of energy consumption level, operating status and potential anomalies; at the same time, combined with the constructed benchmark library, it introduces multi-index difference measurement, dynamic threshold rules and evaluation judgment vector calculation mechanism to complete the quantitative output of intelligent evaluation results of energy-saving facility operation, which has high interpretability and adaptability.

[0078] Therefore, this invention has made systematic improvements in several key aspects such as feature processing, model structure, fusion mechanism and evaluation logic, which significantly improves the automation, accuracy and intelligence of energy-saving facility operation evaluation. Attached Figure Description

[0079] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0080] Figure 1 This is an overall flowchart of an intelligent evaluation method for energy-saving facility operation based on artificial intelligence proposed in this invention;

[0081] Figure 2 This is a schematic diagram of the improved TSMixer model structure of an intelligent evaluation method for energy-saving facility operation based on artificial intelligence proposed in this invention.

[0082] Figure 3 This is a schematic diagram of the dynamic feature fusion network structure of an intelligent evaluation method for energy-saving facility operation based on artificial intelligence proposed in this invention. Detailed Implementation

[0083] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0084] refer to Figure 1-3 An intelligent assessment method for the operation of energy-saving facilities based on artificial intelligence includes the following steps:

[0085] Step 1: Collect multi-source time series data of energy-saving facilities, preprocess the multi-source time series data, and generate context feature vectors;

[0086] Step 2: Extract time step sequences from the multi-source aligned time series, select corresponding context feature vectors for each time step sequence for pairing, and extract supervision labels from the target variable time series according to the preset prediction time range and bind them to the paired samples;

[0087] Step 3: Input the paired samples into the improved TSMixer model, map the time step sequence in the temporal mixing layer, generate feature mixing weight parameters and bias parameters through HyperNetwork with context feature vector as input and adjust the channel dimension mapping relationship in the feature mixing layer, embed the SE-block channel attention structure into the output of the feature mixing layer and combine residual connections and layer normalization to obtain intermediate feature representations;

[0088] Step 4: Repeat Step 3 according to the preset number of layers to obtain the final output features;

[0089] Step 5: Aggregate the final output features along the time dimension and construct a dynamic feature fusion network with the corresponding context feature vectors to form a comprehensive feature set;

[0090] Step 6: Input the comprehensive feature set into the LightGBM model and output the energy consumption prediction results, operating status score and anomaly indication;

[0091] Step 7: Compare the result sequence with the corresponding benchmark data in the benchmark library, and output the intelligent assessment results of energy-saving facility operation.

[0092] In this embodiment, step one specifically includes:

[0093] Multi-source time series data includes equipment operation data, environmental data, and load data;

[0094] Perform timestamp alignment and synchronization on multi-source time series data, and perform missing value processing, noise suppression, and outlier sample removal;

[0095] Units and dimensions are unified, numerical values ​​are normalized and categorically encoded, and time, environmental and operational status features are extracted to generate context feature vectors.

[0096] Determine the data collection objects and data sources, including equipment operation data sources of energy-saving facilities, environmental data sources, and load data sources;

[0097] Set a unified data sampling period and timestamp format, and standardize the timestamps and time zones of the original records from each data source;

[0098] Each data source is resampled according to a set sampling period to generate a time series segment corresponding to the target time axis;

[0099] Remove duplicate records with the same timestamp and sort out-of-order records by timestamp;

[0100] Add device identifiers, channel identifiers, measurement units, and range identifiers to the collected data, and remove records with missing timestamps;

[0101] Align different data sources on a unified timeline to form a multi-source aligned time series that includes device operation data, environmental data, and load data;

[0102] The multi-source aligned time series are archived and indexed according to the data source, device identifier, and channel identifier.

[0103] Establish missing value detection rules, perform forward imputation and time window mean imputation on missing records of continuous fields, perform mode imputation on missing records of discrete fields, and remove records that cannot be imputed.

[0104] Establish a noise suppression process: perform median filtering and moving average smoothing on continuous data, and vote smoothing on discrete data.

[0105] Establish an abnormal sample removal process to screen out abnormal records based on the upper and lower limits of equipment range, engineering constraints, and time change amplitude thresholds.

[0106] The execution units and dimensions are unified, and the measurement units and decimal precision of different data sources are adjusted to the preset standards; numerical normalization processing is performed, and continuous fields are scaled according to preset intervals and standardized with zero mean and unit variance.

[0107] The category field is encoded; one-hot encoding is used for unordered categories, and ordinal encoding is used for ordered categories.

[0108] Extract hour, weekday, weekday and holiday markers from timestamps, extract outdoor temperature, humidity and weather condition markers from environmental data, and extract load level, load change rate and equipment start-up and shutdown markers from load data, and concatenate them in the order of preset fields to form a context feature vector.

[0109] In this embodiment, step two specifically includes:

[0110] Set the window length, sliding step size, and preset prediction time range for the sliding window;

[0111] Extract time step sequences from multi-source aligned time series according to window length and sliding step size to obtain candidate time step sequences;

[0112] For each candidate time step sequence, a corresponding context feature vector is selected according to the timestamp association method to form paired samples. The timestamp association method is used to match the time step sequence with the context feature vector of the corresponding time in multi-source time series data based on the time identifier using the timestamp as an index. If the sampling periods of different data sources are inconsistent, the association is performed according to the principle of closest timestamp or linear interpolation principle.

[0113] Determine the time series of the target variable, and extract the supervision labels corresponding to each paired sample from the time series of the target variable according to the preset prediction time range, and bind the supervision labels to the paired samples;

[0114] The monitoring labels include energy consumption labels, energy efficiency labels, and status labels, which respectively represent the energy consumption, operating efficiency, and operating status of the energy-saving facilities within the predicted time range.

[0115] The determination of the target variable time series includes: selecting candidate monitoring variables from multi-source time series data according to the energy-saving facility operation assessment target; calculating the correlation coefficient between each candidate monitoring variable and the energy consumption benchmark or energy efficiency index; determining the monitoring variables with a correlation coefficient greater than a preset threshold as target variables; and extracting the continuous records of the target variables in the time dimension to generate the target variable time series.

[0116] Candidate samples that do not meet the window integrity requirement will be removed, and candidate samples that cannot obtain supervision labels due to exceeding the preset prediction time range will also be removed.

[0117] Record the start and end timestamps, target variable identifiers, and context feature vector indices for each paired sample to generate a sample set.

[0118] In this embodiment, step three specifically includes:

[0119] Paired samples are input into the improved TSMixer model, wherein the paired samples include time step sequences and corresponding context feature vectors;

[0120] The time-step sequence is mapped in the time-mixing layer and the time-mixing features are output.

[0121] In the feature fusion layer, the context feature vector is used as input to the HyperNetwork to generate feature fusion weight parameters and bias parameters;

[0122] A HyperNetwork submodule is set in the feature mixing layer, which takes the context feature vector as input. Linear mapping, nonlinear activation, and feature compression are performed on the context feature vector to obtain the context hidden layer feature representation. This context hidden layer feature representation is input to a fully connected network with two branches. The first branch outputs a feature mixing weight parameter vector, and the second branch outputs a bias parameter vector. The feature mixing weight parameter vector is dimensionally reshaped to match its shape with the mapping matrix of the temporal mixed features in the channel dimension. The feature mixing weight parameters and bias parameters are passed to the feature mixing layer to dynamically adjust the channel mapping coefficients and bias terms of the temporal mixed features, thereby generating the parameter set of the feature mixing output.

[0123] Based on the feature mixing weight parameters, a linear mapping is performed on the feature variables corresponding to the temporal mixed features at each time step. The weighted summation is performed according to the channel dimension and the bias parameters are superimposed to generate the channel mapping result.

[0124] The channel mapping results are subjected to dimension restoration and shape recovery to obtain feature fusion output;

[0125] For each channel of the feature blending output, an averaging operation is performed in the time dimension. The feature values ​​at each time step are summed and divided by the number of time steps to obtain the average feature value of the current channel. The average feature values ​​of all channels are arranged in order to form the global average pooling result.

[0126] An SE-block channel attention structure is embedded in the feature fusion output, and channel weights are calculated based on the global average pooling results in the time dimension.

[0127] The features of each channel are weighted and recalibrated using channel weights, and intermediate feature representations are obtained by combining residual connections and layer normalization.

[0128] This invention introduces an improved TSMixer model, combining the HyperNetwork dynamic parameter generation mechanism with the SE-block channel attention structure to achieve adaptive fusion modeling of multi-source time series data and contextual features. Specifically, firstly, the contextual feature vector is used as input to the HyperNetwork submodule, and through linear mapping, nonlinear activation, and feature compression, a contextual hidden layer feature representation is generated. Then, a two-branch fully connected network outputs feature mixing weight parameters and bias parameters, and based on these parameters, the channel mapping coefficients and bias terms of the temporal mixed features are dynamically adjusted, thereby achieving adaptive adjustment of feature channel relationships under different operating scenarios. Next, linear mapping and weighted summation are performed on the temporal mixed features to generate channel mapping results, and after dimensionality reduction, a feature mixture output is obtained. Based on this, global average pooling is used to extract channel statistics, and channel weights are calculated in the SE-block attention mechanism to achieve feature weighting and recalibration. Finally, combined with residual connections and layer normalization operations, a structurally stable and semantically rich intermediate feature representation is output, providing high-quality input for subsequent multi-layer feature extraction and evaluation modeling.

[0129] In this embodiment, step four specifically includes: setting a preset number of layers and layer order index, and using the intermediate feature representation as the layer input feature of the first layer; repeating the processing flow described in step three for each layer; using the intermediate feature representation of the current layer as the layer input feature of the next layer and continuing to execute the processing flow described in step three until all preset number of layers are completed; and using the intermediate feature representation of the last layer as the final output feature of the improved TSMixer model.

[0130] In this embodiment, step five specifically includes:

[0131] The final output features are aggregated in the time dimension, and the mean, extreme values ​​and rate of change within the window are calculated to form a time-aggregated representation.

[0132] A dynamic feature fusion network is constructed by taking the temporal aggregation representation and the corresponding context feature vector as input.

[0133] The dynamic feature fusion network includes a random feature aggregation module and a dimension mixing module;

[0134] In the random feature aggregation module, N sub-representations are extracted from the temporal aggregation representation based on a random sampling strategy. During the inference stage, the sub-representations are weighted and converged to generate core features.

[0135] The core features are integrated with the contextual feature vectors to form a fused input;

[0136] In the dimension blending module, a scalable dimension transformation is performed on the fusion input to achieve feature compression and reconstruction, and the output is a fused feature representation;

[0137] Layer normalization and dimensional integration are performed on the fused feature representation to form a comprehensive feature set.

[0138] This invention proposes a dynamic feature fusion network mechanism in the feature fusion stage. By setting a temporal aggregation window, the mean, extreme values, and rate of change of the final output features are statistically calculated to generate a temporal aggregated representation, which, together with the corresponding context feature vector, constitutes the fusion input. The network includes a random feature aggregation module and a dimensionality mixing module. The random feature aggregation module extracts multiple sub-representations from the temporal aggregated representation through a random sampling strategy and performs weighted convergence during the inference stage to generate core features. Subsequently, the core features are integrated with the context feature vector and fed into the dimensionality mixing module, where scalable dimensionality transformation is performed to compress and reconstruct the fused features. Finally, after layer normalization and dimensionality integration, a comprehensive feature set is formed, improving the representational power and adaptability of the fused features.

[0139] In this embodiment, the step of extracting N sub-representations from the temporal aggregation representation based on a random sampling strategy, and then performing weighted convergence on the sub-representations to generate core features during the inference phase, specifically includes:

[0140] First, set the number of samples, random seed, and alignment specifications for the target length and number of channels of the time aggregation representation, and establish an index set for the time dimension and the channel dimension;

[0141] M sets of random masks are generated based on the set number of samplings. Each set of masks includes a time sub-interval index and a channel subset index, which are used to control the time range and feature dimension of each sampling.

[0142] In each sampling process, the corresponding time segment and channel subset are selected from the time aggregation representation based on the random mask, and the sub-segments are extracted as the original sub-representation while maintaining the time order.

[0143] Each time interval and channel subset together constitute a sub-representation;

[0144] Perform length pruning or zero-padding on each sub-representation to keep all sub-representations consistent in time step length, and perform mean removal and amplitude normalization on the channel dimension to align the scale.

[0145] The size-aligned sub-representations are input into the encoding unit, where sub-representation vectors are generated and stored through linear transformations and nonlinear activations.

[0146] During the inference phase, all sub-representation vectors are collected and weights are generated in a preset order to produce input. The fusion weights are calculated for each sub-representation, and the fusion weights are non-negative and normalized to ensure consistent proportions.

[0147] All sub-representations are weighted and summed according to their respective weights to obtain a weighted combination result, which serves as the basic representation of the core feature.

[0148] The weighted results are restored in dimension and shape, mapped to the same tensor structure as the time aggregation representation, and the core features are obtained. At the same time, the sampling index and weight information used are recorded for subsequent fusion and tracing.

[0149] In this embodiment, step six specifically includes:

[0150] Energy consumption labels, operational status rating labels, and anomaly indication labels are extracted from the sample set according to the definition of supervision labels.

[0151] From the time series of the target variables of energy-saving facilities, extract the cumulative energy consumption value or average power value within the corresponding time period according to the preset prediction time range, and use it as an energy consumption label;

[0152] Extract the status parameters (load rate, start-stop ratio, output power fluctuation rate) related to equipment operation from the time series of the target variable, and convert them into operation status score labels through the set score mapping function;

[0153] Based on equipment operation monitoring records or historical operation standards, abnormal intervals (exceeding thresholds, sudden changes, and shutdown events) are marked in the sample set, and abnormal markers for the corresponding time periods are generated as abnormal indication labels.

[0154] The aligned sample set is divided into training set, validation set and test set according to time order and proportion;

[0155] Construct sub-models, including a LightGBM regression sub-model for energy consumption prediction, a LightGBM regression sub-model for operational status scoring, and a LightGBM classification sub-model for anomaly indication;

[0156] Set training parameters for each sub-model, including number of iterations, learning rate, maximum depth, number of leaf nodes, subsample ratio, feature sampling ratio, and regularization coefficient;

[0157] Each sub-model is trained separately on the training set, and the number of iterations and loss convergence are monitored on the validation set to determine the optimal number of iterations.

[0158] Perform inference on the test set for each sub-model corresponding to the optimal number of iterations and output the raw output of energy consumption prediction results, running status score and anomaly indication;

[0159] The original outputs of each sub-model are timestamped according to the sample index to form a result sequence that corresponds one-to-one with the comprehensive feature set;

[0160] The result sequence includes energy consumption prediction results, operating status scores, and anomaly indicators.

[0161] In this embodiment, step seven specifically includes:

[0162] The benchmark records corresponding to the sample set are read from the benchmark library. The benchmark records are indexed by device identifier, device type, load level, environmental conditions and time segment.

[0163] The benchmark library consists of historical operating data, experimental calibration data, and industry standard parameters. The energy consumption benchmark, status scoring benchmark, and anomaly judgment rules in the benchmark library are generated through statistical analysis of historical normal operating condition samples. The benchmark library is stored using a four-level index structure of equipment type, operating condition, environmental conditions, and time period. The benchmark library supports periodic updates and dynamic threshold parameter adaptive adjustment to maintain the matching of evaluation results with the latest operating characteristics.

[0164] The baseline records are retrieved based on the timestamps, device identifiers, and contextual feature vectors in the result sequence, and the baseline data that is consistent with the result sequence in terms of device type, load level, environmental conditions, and time segmentation is selected.

[0165] Align the resulting sequence with the baseline data on the time axis and unify the units of measurement and dimensions; calculate the difference measure and relative difference measure between the energy consumption prediction results and the corresponding baseline energy consumption values ​​to obtain the energy consumption deviation sequence.

[0166] The difference measure and relative difference measure are calculated for the operating status score and the corresponding benchmark score, respectively, to obtain the status score deviation sequence;

[0167] Based on the anomaly indications and the running status rules in the benchmark library, a time-by-time matching is performed to generate an anomaly matching marker sequence;

[0168] Within a preset time window, calculate window statistics for the energy consumption deviation sequence and the state score deviation sequence, and determine dynamic threshold parameters based on threshold rules to determine whether the deviation exceeds the limit within the window.

[0169] The threshold rules are composed of statistical thresholds, operating status rule thresholds, and dynamic adjustment thresholds from the benchmark library. The statistical thresholds are calculated by analyzing the distribution characteristics of historical normal operating condition samples. The operating status rule thresholds are defined according to the operating specifications and industry standards of energy-saving facilities. The dynamic adjustment thresholds are used to correct the static thresholds in real time based on environmental conditions and load change rates.

[0170] The energy consumption deviation sequence, the state score deviation sequence, and the abnormal matching mark sequence are weighted and integrated according to preset weights to form an evaluation judgment vector;

[0171] Based on the assessment judgment vector and threshold rules, intelligent assessment results of energy-saving facility operation are generated. The assessment results include assessment level identifier, out-of-bounds time period, associated equipment identifier and corresponding context feature vector index.

[0172] The assessment results are archived in chronological order and with equipment identification, and the intelligent assessment results of energy-saving facility operation are output.

[0173] Example 1:

[0174] To verify the feasibility of this invention in implementation, it was applied to the operational status assessment of a large group of energy-saving facilities in an industrial park as a test scenario. A comprehensive experiment was conducted to evaluate the proposed AI-based intelligent assessment method for energy-saving facility operation. This park is equipped with various types of energy-saving facilities, including chillers, variable air volume (VAV) air conditioning systems, heat pumps, variable frequency pumps, and fans. The operating environment is complex and significantly affected by external weather, load fluctuations, and personnel density, thus posing a significant assessment challenge.

[0175] In this scenario, 60 days of multi-source time-series data were first collected, with a sampling period of 1 minute, accumulating approximately 86,400 data entries per channel. Data sources included equipment operation data (such as operating current, frequency, temperature, and flow rate), environmental data (such as outdoor temperature and humidity, and solar radiation intensity), and load data (such as cooling load and heating load). The collected raw data underwent preprocessing operations according to the method of this invention, including timestamp alignment, missing value imputation, outlier removal, and unit normalization. Finally, a context feature vector sequence containing 1440 samples per day was constructed.

[0176] Samples were extracted using a sliding window strategy, with each window covering a 60-minute time step and a sliding step size of 10 minutes. The prediction time range was set to the next 30 minutes. Approximately 83,520 paired samples were extracted from the multi-source aligned time series. Supervision labels were extracted from the target variable time series, including actual energy consumption values, state score reference values, and manually labeled anomaly records. After screening and cleaning, 77,800 high-quality samples suitable for modeling were retained.

[0177] The paired samples mentioned above are input into the improved TSMixer model proposed in this invention for training. In the model, the temporal mixing layer effectively extracts the temporal dependencies in the sequence, the feature mixing layer achieves dynamic channel adjustment through a context vector-driven HyperNetwork, and the SE-block structure further enhances the significant differences between channels. Through the superposition of the four layers, the final output is a sequence representation with stronger context awareness.

[0178] A dynamic feature fusion network is applied to deeply integrate output features and context features. In the random feature aggregation module, 10 sub-representations are extracted with a sampling number of 10, and their weighted average is calculated as the core feature. This core feature is then further compressed and reconstructed in the dimension mixing module to finally form a comprehensive feature set.

[0179] The comprehensive feature set was input into three LightGBM sub-models to perform energy consumption prediction, operational status scoring, and anomaly indication, respectively. Model training used 55,000 training samples, 11,000 validation samples, and 11,800 test samples. On the test set, the mean squared error of energy consumption prediction was 0.034 kWh, the mean absolute error between the operational status score and the reference score was 0.12, and the anomaly detection accuracy was 93.6% and the recall rate was 91.8%.

[0180] To evaluate the effectiveness of the assessment mechanism, we constructed a benchmark library, including 150 days of historical operating data, 13 types of experimental calibration data, and standard equipment models, stored using a four-level index structure. By comparing the prediction results with the benchmark data, we can accurately pinpoint the high-energy-consumption operating periods, the frequency of abnormal states, and the key contextual features corresponding to the abnormal operating conditions.

[0181] For example, during the test period from 16:20 to 16:50 on a certain day, this invention identified that the unit energy consumption of the chiller unit was significantly higher than the benchmark value (deviation rate reached 18.3%), accompanied by frequent load fluctuations and drastic changes in external temperature and humidity. The assessment level was marked as "medium-high risk," prompting maintenance personnel to intervene and avoid potential energy waste and equipment damage. Specific experimental data are shown in Table 1:

[0182] Table 1 Summary of Intelligent Assessment Results of Energy-Saving Facility Operation Status

[0183] Equipment Number date Time period Energy consumption deviation rate (%) Status rating bias Anomaly Detection Assessment Level A001 Day 12 16:20 - 16:50 18.3 0.28 yes Medium and high risk A001 Day 13 11:00 - 11:30 5.2 0.05 no normal B103 Day 21 09:10 - 09:40 11.7 0.19 yes Medium risk C052 Day 32 17:40 - 18:10 22.6 0.34 yes High risk A001 Day 42 08:00 - 08:30 3.8 0.02 no normal D089 Day 45 13:30 - 14:00 15.4 0.21 yes Medium and high risk B103 Day 52 10:10 - 10:40 6.9 0.11 no Minor abnormality

[0184] As can be seen from Table 1, the present invention demonstrates significant effects in improving the intelligent perception, assessment accuracy, and timeliness of anomaly warning of the operating status of energy-saving facilities, providing intelligent, efficient, and sustainable technical support for energy-saving operation and maintenance management.

[0185] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent evaluation method for the operation of energy-saving facilities based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Collect multi-source time series data of energy-saving facilities, preprocess the multi-source time series data, and generate context feature vectors; Step 2: Extract time step sequences from the multi-source aligned time series, select corresponding context feature vectors for each time step sequence for pairing, and extract supervision labels from the target variable time series according to the preset prediction time range and bind them to the paired samples; Step 3: Input the paired samples into the improved TSMixer model, map the time step sequence in the temporal mixing layer, generate feature mixing weight parameters and bias parameters with the context feature vector as input in the feature mixing layer through the HyperNetwork submodule and adjust the channel dimension mapping relationship. Embed the SE-block channel attention structure in the output of the feature mixing layer and combine residual connections and layer normalization to obtain the intermediate feature representation. Step 4: Repeat Step 3 according to the preset number of layers to obtain the final output features; Step 5: Aggregate the final output features along the time dimension and construct a dynamic feature fusion network with the corresponding context feature vectors to form a comprehensive feature set; Step 6: Input the comprehensive feature set into the LightGBM model and output the energy consumption prediction results, operating status score and anomaly indication; Step 7: Compare the result sequence with the corresponding benchmark data in the benchmark library, and output the intelligent assessment results of energy-saving facility operation.

2. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 1, characterized in that, Step one specifically includes: Multi-source time series data includes equipment operation data, environmental data, and load data; Perform timestamp alignment and synchronization on multi-source time series data, and perform missing value processing, noise suppression, and outlier sample removal; Units and dimensions are unified, numerical values ​​are normalized and categorically encoded, and time, environmental and operational status features are extracted to generate context feature vectors.

3. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 1, characterized in that, Step two specifically includes: Set the window length, sliding step size, and preset prediction time range for the sliding window; Extract time step sequences from multi-source aligned time series according to window length and sliding step size to obtain candidate time step sequences; For each candidate time step sequence, a corresponding context feature vector is selected according to the timestamp association method to form paired samples; Determine the time series of the target variable, and extract the supervision labels corresponding to each paired sample from the time series of the target variable according to the preset prediction time range, and bind the supervision labels to the paired samples; Candidate samples that do not meet the window integrity requirement will be removed, and candidate samples that cannot obtain supervision labels due to exceeding the preset prediction time range will also be removed. Record the start and end timestamps, target variable identifiers, and context feature vector indices for each paired sample to generate a sample set.

4. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 1, characterized in that, The improved TSMixer model specifically includes: The improved TSMixer model includes a temporal mixing layer, a feature mixing layer, and an SE-block channel attention structure, wherein the feature mixing layer embeds a HyperNetwork submodule; Paired samples are input into the improved TSMixer model, and the paired samples include time step sequences and corresponding context feature vectors; The time-step sequence is mapped in the time-mixing layer and the time-mixing features are output. In the feature fusion layer, the context feature vector is used as input to the HyperNetwork submodule to generate feature fusion weight parameters and bias parameters; A linear mapping is performed to project the context feature vectors into the latent space, followed by nonlinear activation; the context latent layer feature representation is obtained through feature compression. The contextual hidden layer feature representation is fed into two fully connected branches of the HyperNetwork submodule, where the first branch outputs the feature mixing weight parameters and the second branch outputs the bias parameters. Based on the feature mixing weight parameters, a linear mapping is performed on the feature variables corresponding to the temporal mixed features at each time step. The weighted summation is performed according to the channel dimension and the bias parameters are superimposed to generate the channel mapping result. The channel mapping results are subjected to dimension restoration and shape recovery to obtain feature fusion output; For each channel of the feature blending output, an averaging operation is performed in the time dimension. The feature values ​​at each time step are summed and divided by the number of time steps to obtain the average feature value of the current channel. The average feature values ​​of each channel are combined in channel order to form the global average pooling result. An SE-block channel attention structure is embedded in the output of the feature mixing layer, and channel weights are calculated based on the global average pooling results in the time dimension. The intermediate feature representation is obtained by weighting and recalibrating the features of each channel on the output of the feature mixing layer using channel weights, and combining residual connections and layer normalization.

5. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 1, characterized in that, Step four specifically includes: setting a preset number of layers and layer order index, using the intermediate feature representation as the layer input feature of the first layer; repeating the processing flow described in step three for each layer; using the intermediate feature representation of the current layer as the layer input feature of the next layer and continuing to execute the processing flow described in step three until all preset number of layers are completed; using the intermediate feature representation of the last layer as the final output feature of the improved TSMixer model.

6. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 1, characterized in that, Step five specifically includes: The final output features are aggregated in the time dimension, and the mean, extreme values ​​and rate of change within the window are calculated to form a time-aggregated representation. A dynamic feature fusion network is constructed by taking the temporal aggregation representation and the corresponding context feature vector as input. The dynamic feature fusion network includes a random feature aggregation module and a dimension mixing module; In the random feature aggregation module, N sub-representations are extracted from the temporal aggregation representation based on a random sampling strategy. During the inference stage, the sub-representations are weighted and converged to generate core features. The core features are integrated with the contextual feature vectors to form a fused input; In the dimension blending module, a scalable dimension transformation is performed on the fusion input, and the output is a fusion feature representation; Layer normalization and dimensional integration are performed on the fused feature representation to form a comprehensive feature set.

7. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 6, characterized in that, The method of extracting N sub-representations from the temporal aggregate representation based on a random sampling strategy, and performing weighted convergence on the sub-representations to generate core features during the inference phase, specifically includes: First, set the number of samples, random seed, and alignment specifications for the target length and number of channels of the time aggregation representation, and establish an index set for the time dimension and the channel dimension; M sets of random masks are generated based on the set number of samplings. Each set of masks includes a time sub-interval index and a channel subset index. During each sampling process, the corresponding time segment and channel subset are selected from the time aggregation representation based on a random mask; Each time interval and channel subset together constitute a sub-representation; Perform a length pruning operation on each sub-representation to make all sub-representations consistent in time step length, and perform mean removal and amplitude normalization on the channel dimension to align the scale; The size-aligned sub-representations are input into the encoding unit, where sub-representation vectors are generated and stored through linear transformations and nonlinear activations. During the inference phase, all sub-representation vectors are collected and weights are generated in a preset order to form the input. The fusion weights are calculated for each sub-representation, and the fusion weights are nonnegated and normalized. All sub-representations are weighted and summed according to their respective weights to obtain a weighted combination result, which serves as the basic representation of the core feature. The weighted results are then subjected to dimensionality restoration and shape recovery, mapped to the same tensor structure as the time aggregation representation, to obtain the core features.

8. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 1, characterized in that, Step six specifically includes: Energy consumption labels, operational status rating labels, and anomaly indication labels are extracted from the sample set according to the definition of supervision labels. The aligned sample set is divided into training set, validation set and test set according to time order and proportion; Construct sub-models, including a LightGBM regression sub-model for energy consumption prediction, a LightGBM regression sub-model for operational status scoring, and a LightGBM classification sub-model for anomaly indication; Set training parameters for each sub-model, including number of iterations, learning rate, maximum depth, number of leaf nodes, subsample ratio, feature sampling ratio, and regularization coefficient; Each sub-model is trained separately on the training set, and the number of iterations and loss convergence are monitored on the validation set to determine the optimal number of iterations. Perform inference on the test set for each sub-model corresponding to the optimal number of iterations and output the raw output of energy consumption prediction results, running status score and anomaly indication; The original outputs of each sub-model are timestamped according to the sample index to form a result sequence that corresponds one-to-one with the comprehensive feature set; The result sequence includes energy consumption prediction results, operating status scores, and anomaly indicators.

9. The intelligent evaluation method for energy-saving facility operation based on artificial intelligence according to claim 1, characterized in that, Step seven specifically includes: Read the benchmark record corresponding to the sample set from the benchmark library; The baseline records are retrieved based on the timestamps, device identifiers, and contextual feature vectors in the result sequence, and the baseline data that is consistent with the result sequence in terms of device type, load level, environmental conditions, and time segmentation is selected. Align the resulting sequence with the baseline data on the time axis and unify the units of measurement and dimensions; calculate the difference measure and relative difference measure between the energy consumption prediction results and the corresponding baseline energy consumption values ​​to obtain the energy consumption deviation sequence. The difference measure and relative difference measure are calculated for the operating status score and the corresponding benchmark score, respectively, to obtain the status score deviation sequence; Based on the anomaly indications and the running status rules in the benchmark library, a time-by-time matching is performed to generate an anomaly matching marker sequence; Within a preset time window, calculate window statistics for the energy consumption deviation sequence and the state score deviation sequence, and determine dynamic threshold parameters based on threshold rules to determine whether the deviation exceeds the limit within the window. The energy consumption deviation sequence, the state score deviation sequence, and the abnormal matching mark sequence are weighted and integrated according to preset weights to form an evaluation judgment vector; Based on the assessment judgment vector and threshold rules, intelligent assessment results of energy-saving facility operation are generated. The assessment results include assessment level identifier, out-of-bounds time period, associated equipment identifier and corresponding context feature vector index. The assessment results are archived in chronological order and with equipment identification, and the intelligent assessment results of energy-saving facility operation are output.

Images