Air conditioner manufacturing enterprise energy consumption data analysis and prediction method

Abstract

The application discloses an air conditioner manufacturing enterprise energy consumption data analysis and prediction method, including determining the index, range and cycle of air conditioner production energy consumption data; determining five categories of energy consumption data collection boundaries; constructing an energy consumption analysis and prediction feature model; dividing and cross-validating time series data sets; completing initial configuration of the model, and performing iterative verification and hyperparameter tuning, evaluating the model training results; isolating the independent test set which does not participate in training and parameter tuning throughout the process, completing the generalization ability test of the model, forming a feature attribution account and an abnormal error account, and giving an energy consumption model evaluation report; combining real working conditions, completing energy consumption tracing through model prediction, and outputting energy saving optimization control strategies; and then completing the landing operation of the model. The method can realize accurate analysis, prediction and energy saving decision reference of air conditioner production energy consumption data, and provides an effective solution for improving the intelligent level of enterprise energy management.

Description

Technical Field

[0001] This invention relates to the field of air conditioning production data analysis technology, specifically, a method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises. Background Technology

[0002] Currently, energy consumption management in air conditioning manufacturing enterprises mainly relies on manual inspections and traditional threshold alarm methods. This method based on fixed thresholds has the following drawbacks: First, threshold setting depends on experience and cannot quickly adapt to dynamic changes in production load. For example, in scenarios such as full production during peak seasons, reduced production during off-seasons, and process adjustments and switching, the energy consumption baseline values ​​differ significantly, which can easily lead to false alarms or missed alarms. Second, it can only provide post-event alarms when thresholds are exceeded, and cannot predict abnormal energy consumption trends in advance, nor can it help technicians quickly locate the root cause of abnormal energy consumption, thus making it difficult to formulate proactive energy-saving and consumption-reducing strategies.

[0003] Existing methods for controlling energy consumption in air conditioning production lack the ability to integrate and analyze multi-dimensional data from the entire production process. They are unable to construct energy consumption correlation models and real-time dynamic adjustment mechanisms by combining dynamic characteristics such as operating conditions, environment, equipment status, and energy quotas. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method for analyzing and predicting energy consumption data in air conditioning manufacturing enterprises. This method is used to construct an energy consumption analysis model adapted to the dynamic operating conditions of air conditioning production, overcome the limitations of fixed thresholds, and improve the accuracy of anomaly identification. It enables short- and medium-term prediction of energy consumption in air conditioning production, captures abnormal energy consumption trends in advance, and completes the transformation from post-event to pre-event early warning. It also enables the root cause localization of energy consumption anomalies, and combines production data to mine key influencing factors of energy consumption, providing data support for the formulation of energy-saving and consumption-reducing strategies.

[0005] The present invention solves the above problems through the following technical solution:

[0006] A method for analyzing and predicting energy consumption data in air conditioning manufacturing enterprises includes the following steps: determining the indicators, scope, and cycle for collecting, analyzing, and predicting energy consumption data in air conditioning production; defining the collection boundaries for five types of structured and unstructured energy consumption data to avoid data overlap across different dimensions; accessing and cleaning edge-side energy consumption data to prevent data distortion, temporal misalignment, and intrusion of dirty data; extracting features from energy consumption data; constructing an energy consumption analysis and prediction feature model based on the temporal fluctuations and multi-factor coupling characteristics of air conditioning production energy consumption; partitioning and cross-validating the time-series dataset; and completing the initial model configuration based on the energy consumption differences between main and auxiliary equipment in the air conditioning production line. This method aims to achieve a balance between fitting ability and noise-resistant generalization. It uses minimizing the root mean square error (RMSE) of energy consumption prediction on the validation set as the optimization objective, conducting iterative model verification and hyperparameter tuning. Quantitative indicators are used to evaluate the model training results. Independent test sets not involved in training and hyperparameter tuning are isolated throughout the process to test the model's generalization ability, generating feature attribution ledgers and abnormal error ledgers, and providing an interpretable energy consumption model evaluation report. Combined with real-world operating conditions, energy consumption is traced through model prediction, outputting implementable energy-saving optimization and management strategies at the team and workshop levels. The model is deployed near the edge and integrated with multiple business information systems to ensure its successful implementation. This method integrates multi-dimensional data on production, environment, time statistics, and interactive coupling features. Through the construction, training, and evaluation of feature models, it analyzes and predicts energy consumption in intelligent air conditioning production, provides early warnings of anomalies, and formulates energy-saving scheduling optimization strategies.

[0007] As a further improvement of the present invention, the determination of the indicators, range, and cycle for collecting, analyzing, and predicting air conditioning production energy consumption data specifically includes:

[0008] Quantitative monitoring indicators: covering energy consumption accounting indicators of the core air conditioning production workshop, including electricity / water / gas consumption per unit area and comprehensive energy consumption per unit product, eliminating invalid production capacity interference from reworked and repaired non-standard products and test prototypes; energy efficiency indicators of key power equipment, including the full-condition energy efficiency ratio of bending forming machines, fully automatic refrigerant charging machines, whole-machine leak detection equipment, and air compressor units, as well as the energy efficiency indicators of supporting equipment such as chiller units and circulating water pump rooms, to achieve traceability of energy consumption of main and auxiliary equipment;

[0009] Monitoring scope: Achieve four levels of energy consumption monitoring scope: "total company - workshop - production line - key equipment", and link production process points, energy metering points and environmental sensing points to adapt to complex working conditions of staggered production in multiple workshops;

[0010] The operating cycles are tiered: short-term sampling frequencies are 15min / 1h / 4h to meet the dynamic energy consumption monitoring needs of instantaneous equipment cluster start-up and shutdown and process switching in the workshop; medium-term sampling frequencies are 1d / 7d for shift scheduling and monthly rigid energy consumption quotas; long-term sampling frequencies are 30d / 90d for the technical transformation of high-energy-consuming equipment and medium- and long-term energy-saving and carbon reduction planning.

[0011] As a further improvement of the present invention, the step of determining the collection boundaries of five types of structured and unstructured energy consumption data to avoid data overlap of different dimensions specifically includes:

[0012] Basic energy consumption data: real-time / cumulative data from water, electricity, and gas meters. Data time sequence alignment, initial screening, and archiving are completed locally at the edge to prevent data latency and packet loss in the cloud.

[0013] Production operation and planning data: ERP planned output, MES system actual qualified output, standard process rated production cycle time, equipment comprehensive efficiency (OEE) hourly statistics and running time, shift rotation log, start and stop of upstream and downstream process switching, production change shutdown time, annual overhaul shutdown period, and simultaneously mark abnormal working conditions such as temporary shutdown and emergency failure shutdown.

[0014] Environmental data: including workshop temperature and humidity, outdoor temperature and humidity, and atmospheric pressure, used to adapt to the strong correlation between environmental temperature and humidity and the energy consumption of coupled equipment;

[0015] Equipment operation and maintenance data: including speed, power, steam pressure, cooling water flow rate, and maintenance fault records.

[0016] Energy-related data: peak-valley electricity prices and energy consumption allocation for projects.

[0017] As a further improvement of the present invention, the access and cleaning of the edge-side energy consumption data, to prevent data distortion, timing misalignment, and dirty data intrusion, specifically includes the following methods:

[0018] Deploy edge computing power, adopt 5G IIoT acquisition terminals, reuse on-site edge computing gateways, and complete the parsing, time-series normalization, and noise reduction preprocessing of raw energy consumption data at the nearest workshop location. Only transmit structured and effective data to the cloud data acquisition platform to reduce transmission pressure.

[0019] By connecting to the EMS energy management system, MES production execution system, and SCADA equipment monitoring platform through MQTT and OPC UA protocols, the system can achieve access to multi-source heterogeneous data and ledger archiving.

[0020] Hierarchical missing value handling: For instantaneous data loss caused by random instrument offline, interpolation of nearby points before and after the missing data is used to fill in the gaps; for continuous missing data exceeding 3 sampling periods, the missing data is filled in by linking the equipment start / stop log and tracing the process production status; for planned shutdown conditions of equipment with complete power failure, the corresponding time series field is set to "0" and archived; for non-production conditions such as equipment power-on standby and low-load hibernation, the average steady-state standby energy consumption of the same shift, the same time period, and the same equipment over the past 7 days is retrieved for backfilling; for periods marked with production abnormalities, emergency faults, etc., the missing data is directly removed.

[0021] Outlier handling: The "3σ + adaptive box plot" quartile check with a time-series sliding window is adopted instead of a single screening mechanism to adapt to the non-stationary, intermittent pulse-like fluctuation time-series characteristics of energy consumption in air conditioning production, and to eliminate outliers caused by sensor drift and sudden transmission interruptions, but should retain energy consumption pulse fluctuations caused by normal equipment start-up and shutdown and concentrated production.

[0022] Redundant data processing: Based on MD5 global hash encoding, the uniqueness of timestamps and device location codes is verified synchronously to ensure the validity of the modeling dataset in a single time series, single device, and single entry, and to avoid interference from duplicate and redundant data.

[0023] As a further improvement of the present invention, the feature extraction of energy consumption data specifically includes:

[0024] Time feature regularization: Unify the background sampling time benchmark, and extract hourly tags, weekly production schedule tags, monthly production start tags, statutory holiday shutdown tags, and work group shift variable features hourly;

[0025] Nonlinear key feature transformation: Natural logarithmic transformation is applied to nonlinear features; sine-cosine orthogonal periodic encoding is used for strongly periodic time series features to conform to the energy consumption rhythm fluctuation law;

[0026] For feature data that already belongs to the [0,1] interval, it is directly incorporated into the model; while for other cross-level and multidimensional related features, the "Min-Max" algorithm is uniformly used to compress and map them to [0,1].

[0027] As a further improvement of the present invention, the construction of an energy consumption analysis and prediction feature model based on the time-series fluctuations and multi-factor coupling characteristics of air conditioning production energy consumption specifically includes:

[0028] Multi-source fusion benchmark features: integrate core production conditions, environmental features, time-series statistical features and calendar time-series features; process-energy consumption strong correlation interaction features to calculate the correlation coefficient between production cycle and instantaneous energy consumption, the coupling coefficient between equipment start-up time and comprehensive operating energy efficiency, the linkage strength between switching between upstream and downstream processes and instantaneous fluctuations in energy consumption, and superimposed peak-valley time-sharing linkage features.

[0029] Time-series historical lag backtracking features: Construct a time-series energy consumption lag feature sequence, and adaptively determine the optimal gradient energy consumption lag order k based on actual sampling frequency and prediction duration constraints. For short-term predictions of 15 minutes under typical operating conditions, this is to match the inertial fluctuation pattern of instantaneous energy consumption in the workshop;

[0030] Scrolling window steady-state statistical characteristics: A fixed statistical window is used to calculate the extreme values, mean, standard deviation, and time-series slope of energy consumption within the window in real time, so as to draw a visual curve of energy consumption fluctuations and migration over time;

[0031] Time-series differential dynamic characteristics: Calculate the energy consumption difference between the current moment and the previous moment hourly. It can intuitively represent instantaneous changes and pulse fluctuations, and quickly capture the energy consumption changes brought about by the start-up and shutdown of equipment clusters and centralized production.

[0032] As a further improvement of the present invention, the partitioning and cross-validation of the time-series dataset specifically includes:

[0033] The samples are divided into model training set, hyperparameter optimization validation set and independent generalization test set according to time intervals;

[0034] In terms of model validation, an incremental expansion of the segmentation window is adopted instead of random sampling validation to adapt to the real working conditions of rolling iteration on site. At the same time, a time-series cross-validation mechanism is used to preserve the time-series dependencies and conform to the energy consumption pattern of air conditioning production capacity during peak and off-peak seasons.

[0035] As a further improvement of the present invention, the initial configuration of the model is completed based on the energy consumption differences between the main and auxiliary equipment of the air conditioning production line, aiming to achieve a balance between fitting power and noise-resistant generalization, specifically including:

[0036] Learning rate gradient: The initial learning rate baseline range is 0.01~0.1, taking into account both the model's gradient iteration convergence stability and computational limitations;

[0037] Decision tree parameters: The preset number of decision trees n_estimators is in the baseline range of 600~1000, and the number of leaf nodes num_leaves is the default value of 31 for industrial modeling;

[0038] Single leaf parameters: To adapt to different working conditions, the minimum sample size of leaves for core equipment is set to 28~50 to enhance the capture of fine energy consumption features; the minimum sample size of leaves for auxiliary equipment is set to 10~20 to effectively suppress random noise on site.

[0039] To prevent model overfitting, the row dimension sample sampling rate (subsample) and column dimension feature sampling rate (colsample_bytree) are uniformly set to "0.8" in a single round of iterative training.

[0040] And / or, with the goal of minimizing the root mean square error (RMSE) of energy consumption prediction on the validation set, iterative model verification and hyperparameter tuning are performed, specifically including:

[0041] The training set is used to complete the model fitting iteration, the generalization error is monitored through the validation set, and the early stopping mechanism of early_stopping_rounds is enabled.

[0042] Set the number of validation trials n_trials, first roughly screen out the reasonable range of parameters, and then finely search for the optimal hyperparameter combination to reduce the RMSE of the validation set;

[0043] Gradually optimize the number of decision trees n_estimators, the maximum depth of a single tree max_depth, and the minimum number of leaf samples min_child_samples;

[0044] The two core time-series features, the optimal energy consumption lag order k and the rolling statistical window size, are simultaneously incorporated into the hyperparameter search space to achieve mutual synergy between model structure and feature time series.

[0045] As a further improvement of the present invention, the step of combining real working conditions and using model prediction to complete energy consumption source tracing, and outputting energy-saving optimization and management strategies that can be implemented at the team and workshop levels, specifically includes:

[0046] By combining the importance ranking of the model's original features with the ranking of SHAP values, a process-energy consumption correlation ledger is established, and typical thresholds are identified.

[0047] The system plots a time-series comparison curve of predicted and actual energy consumption values, identifies abnormal deviations from the warning threshold for energy consumption, and locates inefficient equipment and high-energy-consuming processes. It inputs monthly production capacity fluctuations and seasonal temperature change ranges, outputs the predicted energy consumption probability distribution, provides risk assessment, and completes dynamic load scheduling.

[0048] Develop peak-valley staggered production scheduling plans and provide optimization suggestions for the staggered start-up and shutdown of high-energy-consuming equipment; implement zoned collaborative control of temperature and humidity for workshop process management.

[0049] As a further improvement of the present invention, the method also includes: establishing a full life cycle model monitoring, gradient iterative optimization, and version backtracking to ensure the stability and long-term application of annual energy consumption prediction.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] This method integrates multi-dimensional data on production, environment, time statistics, and interactive coupling features. Through the construction, training, and evaluation of feature models, it analyzes and predicts intelligent air conditioning production energy consumption, provides early warnings of anomalies, and formulates energy-saving scheduling optimization strategies. Furthermore, based on the energy consumption control requirements of air conditioning production, this method matches four levels of energy consumption data with three cyclical periods, matching them with business scenarios such as enterprise energy scheduling, monthly and annual production planning, and energy-saving planning. Combined with practical experience, it summarizes a method for energy consumption data analysis and prediction in air conditioning enterprises. This method enables accurate analysis and prediction of air conditioning production energy consumption data and provides reference for energy-saving decisions, offering an effective solution for improving the intelligent level of enterprise energy management. Attached Figure Description

[0052] Figure 1 This is a flowchart of a method for analyzing and predicting energy consumption data in air conditioning production according to the present invention;

[0053] Figure 2 This is a diagram illustrating the architecture of the energy consumption analysis and prediction feature model of this invention. Detailed Implementation

[0054] 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 only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1:

[0056] See attached document Figure 1-2 This embodiment provides a method for analyzing and predicting energy consumption data in air conditioning production, the specific steps of which include:

[0057] S110. Determine the indicators, scope, and cycle for collecting, analyzing, and predicting energy consumption data in air conditioning production; including:

[0058] 1.1.1 Quantitative monitoring indicators: covering energy consumption accounting indicators for core air conditioning production workshops such as sheet metal stamping, plastic injection molding, complete unit assembly, and performance testing. Energy consumption accounting indicators include electricity / water / gas consumption per unit area and comprehensive energy consumption per unit product (SEC). Comprehensive energy consumption per unit product = workshop equivalent comprehensive energy consumption (standard coal equivalent kgce) within the period ÷ qualified complete unit output within the period, excluding ineffective production capacity interference such as rework, repair of non-standard products, and test prototypes.

[0059] Key power equipment energy efficiency indicators, including bending forming machines, fully automatic refrigerant filling machines, whole machine leak detection equipment, full-condition energy efficiency ratio of air compressor units, and energy efficiency indicators of supporting equipment such as chiller units and circulating water pump rooms, to achieve traceability of energy consumption of main and auxiliary equipment.

[0060] 1.1.2 Monitoring Level and Scope: Achieve four levels of energy consumption monitoring scope: "Company Total - Workshop - Production Line - Key Equipment", and connect production process points, energy metering points, and environmental sensing points to adapt to complex working conditions of staggered production in multiple workshops.

[0061] 1.1.3. Differentiated operating cycles: Short-term sampling frequencies are 15min / 1h / 4h to meet the dynamic energy consumption monitoring needs of instantaneous equipment cluster start-up and shutdown and process switching in the workshop; medium-term sampling frequencies are 1d / 7d for shift scheduling and monthly rigid energy consumption quotas; long-term sampling frequencies are 30d / 90d for the technical transformation of high-energy-consuming equipment and medium- and long-term energy conservation and carbon reduction planning.

[0062] S120. Determine the collection boundaries for five types of structured and unstructured energy consumption data to avoid data overlap across different dimensions; including:

[0063] 1.2.1 Energy consumption basic data: real-time / cumulative data of water, electricity and gas meters, collected every 5 / 15 minutes. Data time sequence alignment, initial screening and archiving are completed locally on the edge side to prevent data delay and packet loss in the cloud.

[0064] 1.2.2 Production Operation and Planning Data: ERP planned output, MES system actual qualified product output, standard process rated production cycle time, equipment comprehensive efficiency (OEE) hourly statistics and runtime, shift rotation log, start and stop of upstream and downstream processes, production changeover downtime, annual overhaul downtime, and simultaneously mark abnormal operating conditions such as temporary shutdown and emergency failure shutdown.

[0065] 1.2.3 Environmental data: including workshop temperature and humidity, outdoor temperature and humidity, and atmospheric pressure, used to adapt to the strong correlation between environmental temperature and humidity and equipment energy consumption.

[0066] 1.2.4 Equipment operation and maintenance data: speed, power, steam pressure, cooling water flow, maintenance fault records, etc.

[0067] 1.2.5 Energy Support Data: Peak-valley electricity prices, energy consumption allocation for projects, etc.

[0068] S130. Access and cleaning of edge-side energy consumption data to prevent data distortion, timing misalignment, and intrusion of dirty data; including:

[0069] 1.3.1 Deploy edge computing power, adopt 5G IIoT (Industrial Internet of Things) acquisition terminals, reuse on-site edge computing gateways, and complete the parsing, time-series normalization, and noise reduction preprocessing of raw energy consumption data at the nearest workshop location. Only transmit structured and valid data to the cloud data acquisition platform to reduce transmission pressure.

[0070] 1.3.2. Connect to the EMS energy management system, MES production execution system, and SCADA equipment monitoring platform via MQTT and OPC UA protocols to achieve access to multi-source heterogeneous data and ledger archiving.

[0071] 1.3.3 Hierarchical missing value handling: For instantaneous data loss caused by random instrument offline, interpolation of nearby points before and after the missing data is used to fill the gap; for continuous missing data exceeding 3 sampling periods, the missing data is filled by linking the equipment start / stop log and tracing the process production status; for planned shutdown conditions of equipment with complete power failure, the corresponding time series field is set to "0" and archived; for non-production conditions such as equipment power-on standby and low-load hibernation, the average steady-state standby energy consumption of the same shift, the same period, and the same equipment in the past 7 days is retrieved for backfilling; for periods marked with production abnormalities, emergency faults, etc., the missing data is directly removed.

[0072] 1.3.4 Outlier Handling: The "3σ + Adaptive Box Plot" quartile verification with a time-series sliding window is adopted instead of a single screening mechanism to adapt to the non-stationary, intermittent pulse-like fluctuation time-series characteristics of energy consumption in air conditioning production, and to eliminate outliers caused by sensor drift and sudden transmission interruptions, but should retain energy consumption pulse fluctuations caused by normal equipment start-up and shutdown and concentrated production.

[0073] 1.3.5 Redundant Data Processing: Based on MD5 global hash encoding, the uniqueness of timestamps and device location codes is verified synchronously to ensure the validity of the modeling dataset in a single time series, single device, and single entry, and to avoid interference from duplicate and redundant data.

[0074] S140. Extract features from energy consumption data:

[0075] 1.4.1 Time Feature Regularization: Unify the background sampling time benchmark, and extract hourly labels, weekly production schedule labels, monthly production start-up labels, statutory holiday shutdown labels, and shift variable features hourly to ensure strict homogeneity of the time sequence of the entire dataset.

[0076] 1.4.2 Nonlinear Key Feature Transformation: Natural logarithmic transformation is performed on nonlinear features such as actual output in the workshop and instantaneous load rate of equipment; sine-cosine orthogonal periodic coding is adopted for strongly periodic time series features such as intraday hourly and minute rhythms to conform to the energy consumption rhythm fluctuation law.

[0077] 1.4.3 For feature data such as equipment load rate and effective process operating rate that belong to the [0,1] interval, they can be directly incorporated into the model; while other cross-level and different dimension correlation features are uniformly compressed and mapped to [0,1] using the "Min-Max" algorithm to eliminate the influence of data with different dimensions and improve the efficiency of iterative training to accelerate model convergence.

[0078] S150. Based on the temporal fluctuations and multi-factor coupling characteristics of air conditioning production energy consumption, an energy consumption analysis and prediction feature model is constructed:

[0079] 1.5.1 Multi-source fusion benchmark features: Integrating core production conditions (real-time qualified product output, shift rotation, equipment start-up and shutdown status), environmental features (dry and wet bulb temperature and humidity data at key points in the workshop), time-series statistical features (extreme values, mean values, and standard deviations of energy consumption per window), calendar time-series features (working days / shutdown days, seasonal change sequence, and intraday time-of-use energy consumption features); strong correlation and interaction features between process and energy consumption, to calculate the correlation coefficient between production cycle and instantaneous energy consumption, the coupling coefficient between equipment start-up time and comprehensive operating energy efficiency, the linkage strength between switching between upstream and downstream processes and instantaneous fluctuations in energy consumption, and superimposed features such as peak-valley time-of-use linkage.

[0080] 1.5.2 Time-series historical lag backtracking features: Construct energy consumption time-series lag feature sequences energy_lag_1, energy_lag_2 to energy_lag_k, and adaptively determine the optimal gradient energy consumption lag order k based on the actual sampling frequency and prediction duration constraints. For short-term predictions of 15 minutes under typical operating conditions, k can be selected from 1 to 6 to match the inertial fluctuation pattern of instantaneous energy consumption in the workshop.

[0081] 1.5.3 Scrolling Window Steady-State Statistical Characteristics: Fix 1h, 4h, and 24h statistical windows, and calculate the extreme values, mean, standard deviation, and time-series change slope of energy consumption within the window in real time to draw visualization curves of energy consumption fluctuations and migration over time.

[0082] 1.5.4. Time-series differential dynamic characteristics: Calculate the energy consumption difference between the current moment and the previous moment hourly. It can intuitively represent instantaneous changes and pulse fluctuations, and quickly capture the energy consumption changes brought about by the start-up and shutdown of equipment clusters and centralized production.

[0083] S160. Partitioning and cross-validating the time-series dataset:

[0084] 1.6.1. The earliest 80% of the time series samples are designated as the model training set, the middle 10% of the samples are designated as the hyperparameter optimization validation set, and the latest 10% of the samples are designated as the independent generalization test set.

[0085] 1.6.2 In terms of model verification, an incremental expansion of the segmentation window is adopted instead of random sampling verification to adapt to the real working conditions of rolling iteration on site. At the same time, a time-series cross-validation mechanism is used to retain the time-series dependencies and conform to the energy consumption pattern of air conditioning production capacity during peak and off-peak seasons.

[0086] S170. Based on the energy consumption differences between the main and auxiliary equipment in the air conditioning production line, complete the initial model configuration, and strive for a balance between fitting power and noise-resistant generalization:

[0087] 1.7.1 Learning Rate Gradient: The initial learning rate is within the range of 0.01 to 0.1, balancing the model's gradient iteration convergence stability with computational limitations. A small learning rate can prevent overfitting and is suitable for modeling the energy consumption of key core equipment, but the smaller the rate, the slower the convergence iteration. A large learning rate is suitable for extrapolating the total energy consumption of the workshop and can improve convergence efficiency.

[0088] 1.7.2 Decision Tree Parameters: The preset number of decision trees n_estimators is in the baseline range of 600~1000, and the number of leaf nodes num_leaves is the default value of 31 for industrial modeling. This can avoid the problem of excessive computing power consumption due to excessive model size or insufficient fitting power due to excessively small model size.

[0089] 1.7.3 Single Leaf Parameters: To adapt to different working conditions, the minimum sample size of leaf (min_data_in_leaf) for core equipment (bending machine, refrigerant charging machine, fully automatic leak detection equipment, etc.) is set to 28~50 to enhance the capture of refined energy consumption characteristics; the minimum sample size of leaf for auxiliary equipment (air compressor unit, circulating cooling tower, etc.) is set to 10~20 to effectively suppress random noise on site.

[0090] 1.7.4 To prevent model overfitting, the row dimension sample sampling rate (subsample) and column dimension feature sampling rate (colsample_bytree) in a single round of iterative training are uniformly set to "0.8" to avoid the risk of overfitting in a single run.

[0091] S180. With minimizing the root mean square error (RMSE) of energy consumption prediction on the validation set as the optimization objective, iteratively verify the model and fine-tune the hyperparameters:

[0092] 1.8.1 Use the training set to complete the model fitting iteration, monitor the generalization error change through the validation set, and enable the early stopping mechanism of early_stopping_rounds=50, that is, if the validation error has not been optimized for 50 consecutive rounds, then the training will be terminated.

[0093] 1.8.2 The number of validation trials, n_trials, is set to 100~200. First, a reasonable range of parameters is roughly screened out, and then the optimal hyperparameter combination is searched in detail to reduce the RMSE of the validation set.

[0094] 1.8.3 Gradually optimize the number of decision trees (n_estimators), the maximum depth of a single tree (max_depth), and the minimum number of leaf samples (min_child_samples) to improve the model's generalization ability across different working conditions and time periods.

[0095] 1.8.4 Incorporate the two core time-series features, namely the optimal energy consumption lag order k and the rolling statistical window size, into the hyperparameter search space to achieve mutual synergy between model structure and feature time series.

[0096] S190. The following metrics are used to evaluate the model training results:

[0097] 1.9.1 Quantitative indicators include RMSE (Root Mean Square Error of Energy Consumption Prediction), MAE (Mean Absolute Deviation), WMAPE (Weighted Absolute Percentage Error of Measured Energy Consumption), and R² (Coefficient of Determination). WMAPE uses the hourly actual energy consumption as a weighted benchmark, preventing interference from amplifying small errors during low-load standby periods. A threshold of R² ≥ 0.85 is set, with an ideal target of R² ≥ 0.9. The closer R² is to 1, the stronger the model's fit. If R² is below 0.85, the model's rationality, hyperparameter suitability, and data source accuracy should be considered, requiring iterative optimization.

[0098] 1.9.2. Based on the marginal contribution of each characteristic to energy consumption, sort by global average absolute SHAP value to quantify the contribution of production capacity, ambient temperature and humidity, equipment operating time, etc. to energy consumption fluctuations, identify high energy consumption sources, and provide a reference for technological transformation.

[0099] S210. Isolate the independent test set that did not participate in training and parameter tuning throughout the process, complete the generalization ability test of the model, form a feature attribution ledger and anomaly error ledger, and give an interpretable energy consumption model evaluation report.

[0100] S211. Based on real-world working conditions, use model predictions to trace energy consumption sources and output energy-saving optimization and management strategies that can be implemented at the team and workshop levels:

[0101] 2.11.1. Combining the importance ranking of the model's original features with the ranking of SHAP values, establish a process-energy consumption correlation ledger and identify typical thresholds, such as the power consumption increment corresponding to an increase of 10 units per hour in overall machine capacity, the energy consumption increment of the air compressor unit corresponding to an increase of 1°C in ambient temperature, etc.

[0102] 2.11.2. Plot the time-series comparison curve of predicted and actual energy consumption values ​​to identify abnormal deviations from the warning threshold and locate inefficient equipment and high-energy-consuming processes. Input monthly production capacity fluctuations and seasonal temperature variation ranges, output the predicted energy consumption probability distribution, provide risk assessment, and complete dynamic load scheduling.

[0103] 2.11.3. Develop peak-valley staggered production scheduling plan and provide optimization suggestions for the staggered start-up and shutdown of high-energy-consuming equipment; implement zoned collaborative control of temperature and humidity for workshop process management.

[0104] S212. The model is deployed and run by using edge-based localization and multi-service information system integration.

[0105] The system adopts a parallel approach of "offline + online". The offline background inference automatically reviews historical energy consumption during the daily off-peak computing power period and predicts the energy consumption baseline for the next day. The online real-time inference is encapsulated through API and pushes the energy consumption prediction results.

[0106] The model uses serialized encrypted storage to prevent data and energy consumption ledger leakage, and completes the visualization of energy efficiency analysis, energy consumption prediction and anomaly alarm on PC and mobile terminals.

[0107] S213. Establish full lifecycle model monitoring, gradient iterative optimization, and version rollback to ensure stable and controlled annual energy consumption forecasting and long-term application:

[0108] 2.13.1 Complete the intelligent early warning of instantaneous energy consumption anomalies. If the measured energy consumption deviates from the model prediction baseline by ±10%, a graded audible and visual alarm will be triggered. Link the intelligent production scheduling system to dynamically optimize staggered production and flexible load allocation energy management.

[0109] 2.13.2 The backend continuously monitors RMSE, MAE, R², and MAPE metrics 24 / 7, and immediately pushes maintenance work orders when an anomaly is triggered.

[0110] 2.13.3. Based on the steady-state model prediction performance over the past 90 days, if the overall prediction accuracy decreases by more than 20% and the steady state persists for 15 consecutive days, or if there are major changes in operating conditions such as process upgrades, replacement of core equipment, or technical modifications to core processes, the model will be retrained and optimized.

[0111] 2.13.4 Monthly small sample incremental fine-tuning is performed to adapt to minor fluctuations in operating conditions; quarterly data retraining is performed to adapt to seasonal energy consumption changes; and a model version ledger and a one-click rollback mechanism for abnormal versions are used to ensure that the average absolute percentage error (MAPE) of energy consumption prediction for the whole year is less than 10%.

[0112] Example 2:

[0113] This embodiment focuses on the core production equipment of an air conditioning manufacturing company's sheet metal workshop, final assembly workshop, injection molding workshop, and performance testing workshop, along with auxiliary production facilities such as refrigerant stations and air compressor stations. It details the specific implementation steps of a method for analyzing and predicting air conditioning production energy consumption data:

[0114] 1. Determine the indicators, scope, and cycle for collecting, analyzing, and predicting energy consumption data in air conditioning production:

[0115] 1.1) Based on the company's production process characteristics, the following energy consumption analysis and forecast indicators are defined:

[0116] (1) Electricity consumption per unit area (kWh / ㎡·d), water consumption (m³ / ㎡·d), gas consumption (Nm³ / ㎡·d) and comprehensive energy consumption (kgce / ㎡·d) of each workshop.

[0117] (2) The average daily comprehensive energy consumption at the company level and the average daily comprehensive energy consumption at the workshop level are both expressed in kgce;

[0118] (3) Unit product comprehensive energy consumption (SEC), SEC = workshop daily comprehensive energy consumption (kgce) ÷ daily output of qualified products entering the warehouse (pieces);

[0119] (4) Energy efficiency ratio of key equipment, of which bending machine, refrigerant charging machine and leak detection equipment are based on unit product power consumption (kWh / piece); air compressor unit is based on unit gas production power consumption (kWh / m³) to characterize its energy utilization efficiency.

[0120] 1.2) Construct a four-level energy consumption monitoring scope:

[0121] (1) Enterprise level: Covers the production workshops and auxiliary facilities (refrigerant station, air compressor station, warehouse) in the park, and monitors the comprehensive energy consumption, energy structure and energy efficiency level of enterprises.

[0122] (2) Workshop level: Covering the core workshops of sheet metal, final assembly, injection molding and performance testing, monitoring the comprehensive energy consumption, energy consumption per unit area and energy consumption per unit product of each workshop.

[0123] (3) Production line level: Monitor the energy consumption, production cycle and energy consumption matching degree of each production line.

[0124] (4) Critical equipment level: Monitor the energy consumption, coefficient of performance (COP), and operating status of a single piece of equipment.

[0125] 1.3) Based on the enterprise's production plan, scheduling needs, and energy management requirements, define the three-level monitoring and forecasting cycle:

[0126] (1) Short-term: 15min / 1h / 4h, of which the 15min frequency is used for real-time monitoring and scheduling of energy consumption of key equipment (such as air compressor start-up and shutdown adjustment), the 1h frequency is used for real-time monitoring of production line energy consumption, and the 4h frequency is used for real-time scheduling of workshop energy consumption. A short-term energy consumption scheduling report is generated daily for the optimization of production energy consumption on the same day.

[0127] (2) Mid-term: 1d / 7d, daily statistics of energy consumption data of each workshop and production line, comparison of planned energy consumption with actual energy consumption; weekly energy consumption analysis report, optimization of production schedule and energy consumption plan for the following week.

[0128] (3) Long-term: 30d / 90d, monthly statistics of energy consumption data of enterprises and workshops, and calculation of energy consumption per unit product; quarterly energy consumption trend analysis, in summer for high temperature and high humidity conditions, optimize the matching of air compressor and refrigeration load, and in winter for the insulation and gas efficiency of injection molding workshop to formulate energy-saving adjustment plan.

[0129] 2. Based on the company's existing information systems and actual production situation, clarify the following data collection scope and dimensions:

[0130] 2.1) Basic Energy Consumption Data:

[0131] (1) Power data: workshop main electricity meter, production line sub-meters, key equipment independent electricity meters, collection frequency 15min, data accuracy 0.01kWh.

[0132] (2) Water resources data: the main water meter in the workshop and the water meter of key processes, with a collection frequency of 15 minutes and a data accuracy of 0.01 m³.

[0133] (3) Gas data: Gas flow meter in the injection molding workshop, with a collection frequency of 15 min. The original data unit is Nm³, which is converted to standard coal equivalent kgce according to national standards, with a conversion accuracy of 0.01 kgce.

[0134] 2.2) Production planning and operational data:

[0135] (1) Production data: The planned daily output was 3,100 sets, and the actual daily output was 3,012 sets; the planned hourly output of production line No. 1 in the general assembly workshop was 200 sets, and the actual hourly output was 192 sets.

[0136] (2) Production cycle time: The average time per unit in the final assembly workshop is 4.2 min / set, and the actual cycle time is 4.4 min / set; the standard cycle time in the injection molding workshop is 2.9 min / piece, and the actual cycle time is 2.9 min / piece.

[0137] (3) Other data: Equipment utilization rate is 88.5% (bending machine) and 91.2% (refrigerant charging machine); the average operating time of key equipment is 10.5h / d; a three-shift system is implemented, the average process switching time is 27min / time, the planned production changeover time is 45min / time, and the planned downtime is 12:00~12:30 and 20:00~20:15 every day (equipment inspection).

[0138] 2.3) Environmental data: Environmental monitoring points were set up in each workshop and factory area, with data collection frequency of 15 minutes.

[0139] (1) Workshop temperature and humidity: Sheet metal workshop 25.7℃, 62%RH; final assembly workshop 24.6℃, 58%RH; injection molding workshop 28.7℃, 55%RH; testing workshop 23.5℃, 60%RH.

[0140] (2) Outdoor temperature and humidity: 26.5℃, 65%RH; atmospheric pressure 101.2kPa.

[0141] 2.4) Collect key equipment operation and maintenance parameters to support the tracing of energy consumption anomalies:

[0142] (1) Maintenance records: The last maintenance time for the bending machine was the 7th of this month, and the maintenance content was lubricating oil replacement and precision calibration; the last maintenance time for the refrigerant charging machine was the 11th of this month.

[0143] (2) Fault record: The No. 2 air compressor unit experienced a pressure abnormality fault on a certain working day last month. It was shut down for maintenance for 2.8 hours, and the energy consumption during the fault period was 0.

[0144] (3) Operating parameters: bending machine speed 1450r / min, power 4.12kW; refrigerant filling machine filling pressure 0.45MPa; air compressor cooling water flow rate 12m³ / h.

[0145] 2.5) Energy data: Peak electricity price in spring and autumn is 1.1285 yuan / kWh, normal price is 0.7412 yuan / kWh, and off-peak price is 0.3539 yuan / kWh.

[0146] 3. Collect and clean energy consumption data:

[0147] 3.1) Deploy 5G edge computing IIoT acquisition terminals in each workshop to preprocess energy consumption data locally using edge computing, remove obvious outliers, and standardize the data format. For example, if the power of the refrigerant charging machine abnormally jumps from about 4.0kW to 20.8kW at a certain moment, the system identifies this as abnormal data at the edge and removes the outlier, preventing it from being uploaded to the cloud.

[0148] 3.2) Multi-system data integration:

[0149] (1) Connect to the energy management system (EMS) via OPC UA to obtain the cumulative energy consumption data of water, electricity and gas, with a connection delay of ≤30s.

[0150] (2) Connect to the Manufacturing Execution System (MES) via MQTT to obtain production data such as output, production cycle time, and shifts. The connection delay is ≤1min.

[0151] (3) Connect to the SCADA monitoring system to obtain key equipment operating parameters (speed, power, pressure) and environmental data. The connection delay is ≤15s.

[0152] (4) Connect to the equipment operation and maintenance system to obtain maintenance and fault records, realize unified access of multi-source data, and achieve an effective data integration rate of ≥98%.

[0153] 3.3) Handling missing values:

[0154] (1) Random missing data: The power consumption data of production line 1 in the final assembly workshop from 10:30 to 10:45 is missing. The linear interpolation method is used. Based on the data from 10:15 to 10:30 (128.63 kWh) and 10:45 to 11:00 (132.45 kWh), the power consumption of the missing period is 130.54 kWh.

[0155] (2) Continuous missing data: The gas data of the injection molding workshop was missing for three consecutive 15-minute periods. Based on the equipment status (normal operation), the average value of adjacent periods (18.66 kgce, 18.83 kgce, 18.79 kgce) was used to fill the missing data. The data after filling the missing data was 18.76 kgce / period.

[0156] (3) Shutdown status: The air compressor is shut down for 2.5 hours due to a fault. The energy consumption data for the corresponding time period is set to "0". There are a total of 10 15-minute time periods, and the energy consumption is recorded as "0" for each period.

[0157] (4) Standby status: The bending machine is in standby mode at night (24:00~8:00). The average standby energy consumption of the past 7 days is used to fill the 32 time periods, and each time period is filled with 0.32kWh.

[0158] (5) Due to equipment failure, the production in the final assembly workshop was interrupted from 10:10 to 15:20. During this period, a total of 20 15-minute cycle energy consumption data were uniformly marked and removed, and were not included in the modeling.

[0159] 3.4) Outlier handling: Outliers are identified using a combination of the "3σ principle" and the box method.

[0160] (1) Collect the electricity consumption data of the final assembly workshop for the previous month. The mean was 25200 kWh, the standard deviation σ was 1250 kWh, and the 3σ range was 21450~28950 kWh. The 3σ range was used to identify anomalies in the daily electricity consumption. For fine-grained 15-minute electricity consumption data, the box method was used for identification. The interquartile range (IQR) was calculated to be 1800 kWh, and the anomaly threshold was 18900~31500 kWh.

[0161] (2) A total of 2,880 15-minute power consumption data points over 30 days were jointly identified, and 11 abnormal data points were removed.

[0162] 3.5) Redundant data processing:

[0163] Based on the MD5 hash deduplication algorithm, an MD5 hash value is calculated for each energy consumption record. Records with the same hash value are considered duplicates, and only one record is retained. This process deduplicates the collected duplicate data. For example, if three identical workshop electricity consumption samples (128.63 kWh) are collected by the IIoT terminal and EMS system, they are identified as duplicate records by MD5 hashing, and only one record is retained, thus reducing data redundancy.

[0164] 4. Extract features from the cleaned data:

[0165] 4.1) Regularized time characteristics:

[0166] (1) Extract time features: hour, weekday, month, holidays (weekday = 0, holidays = 1), shift (morning, noon and evening are 1 / 2 / 3 respectively).

[0167] (2) Unified sampling frequency: The equipment operating parameters (5-minute acquisition cycle), environmental data (15-minute), and energy consumption data (5 / 15-minute) are uniformly resampled to 15 minutes. For example, the 5-minute power data of the bending machine (4.06kW, 4.12kW, 4.07kW) are resampled to a 15-minute average of 4.08kW to align with the timing nodes.

[0168] 4.2) Feature transformation;

[0169] (1) Logarithmic transformation of nonlinear characteristics: Logarithmic transformation of non-negative and nonlinear production characteristic values ​​can reduce the impact of nonlinearity on the model. For example, the hourly output of the final assembly workshop is 172 sets, which can be transformed into ln(172)≈5.15.

[0170] (2) Periodic feature encoding: Sine or cosine periodic encoding is performed on the hourly features to preserve periodic feature information. For example, 10:20 is encoded as sin(2π*((10+20 / 60) / 24))≈0.4226.

[0171] 4.3) Normalization process:

[0172] (1) No need to normalize data: The equipment load rate and utilization rate are in the range of [0,1], so the original data is directly retained.

[0173] (2) For data such as power consumption, water consumption, temperature, and humidity that are not in [0,1], Min-Max normalization is used to map them to [0,1].

[0174] x_norm=(x-x_min) / (x_max-x_min)

[0175] If the power consumption (kWh) in the final assembly workshop is in the range of [100, 150] / 15min, then “128.63” is normalized to “0.5726”.

[0176] 5. Construct a characteristic model for energy consumption analysis and prediction in air conditioning production:

[0177] 5.1) Integrate various features to construct a multi-feature coupled feature set:

[0178] (1) Basic characteristics: Production characteristics such as output, shift, equipment status, load rate; Environmental characteristics such as workshop temperature and humidity; Time-series statistical characteristics: 15-minute power consumption maximum value 132.45kWh, minimum value 128.63kWh, mean 130.54kWh, standard deviation 1.91kWh; Time characteristics such as weekday, March, 10 o'clock.

[0179] (2) Interaction characteristics: Production cycle time-energy consumption interaction characteristics, such as the time lag correlation coefficient between production cycle time of 3.4 min / set and power consumption of 130.54kWh is 0.82; Start-up time-energy efficiency interaction characteristics, such as the correlation coefficient between start-up time of 10.5 h / d and COP4.2 is 0.78; Process switching-energy consumption fluctuation, such as the energy consumption fluctuation amplitude corresponding to process switching time of 28 min / time is 12.3%.

[0180] (3) Linkage characteristics: The linkage characteristics of public works in the park, such as peak and valley electricity prices and energy consumption sharing of 33% in the final assembly workshop.

[0181] 5.2) Determine the historical energy consumption lag order k by combining the prediction period and sampling frequency:

[0182] (1) Short-term prediction (15min): k=1~6, that is, extract the energy consumption lag values ​​of the first 15min (energy_lag_1=128.63 kWh), the first 2 15min (energy_lag_2=127.85 kWh)..., the first 6 15min (energy_lag_6=126.32 kWh) and incorporate them into the feature model.

[0183] (2) Mid-term forecast (1d): k=1~24, extract the energy consumption lag value of each 15min period of the previous day.

[0184] 5.3) Calculate the statistical characteristics of different scrolling windows:

[0185] (1) In the past hour, there were four 15-minute periods: the average energy consumption was 129.87 kWh, the standard deviation was 1.76 kWh, and the slope was 0.23. The energy consumption showed a slight upward trend.

[0186] (2) Over the past 4 hours (16 15-minute intervals): the average energy consumption was 130.25 kWh, the standard deviation was 2.18 kWh, and the slope was 0.15, indicating that the energy consumption was basically stable.

[0187] (3) Over the past 24 hours (96 15-minute intervals): the average energy consumption was 128.93 kWh, the standard deviation was 2.85 kWh, and the slope was -0.08, showing a slight downward trend in energy consumption.

[0188] 5.4) Calculate the energy consumption difference between the current time and the previous time to reflect the rate of energy consumption change. For example, the power consumption in the final assembly workshop from 10:00 to 10:15 is 128.63 kWh, while the power consumption at the previous time (9:45 to 10:00) is 127.85 kWh. The difference is 0.78 kWh, indicating a slight increase in energy consumption at a rate of 0.78 kWh / 15 min.

[0189] 6. Dataset partitioning:

[0190] 6.1) Data set proportions: the earliest 80% of the time series data is used as the training set (the first 4.8 months, 144 days, a total of 13,824 records), the next 10% is used as the validation set (0.6 months, 18 days, a total of 1,728 records), and the last 10% is used as the test set (0.6 months, 18 days, 1,728 records), strictly divided according to the time order.

[0191] 6.2) An incremental window partitioning method is adopted. The initial training window is 30 days. Subsequently, one day's data is added to the training set each time the window is rolled over. The validation set is kept to a single day until the entire training interval is covered, thus completing the rolling validation. This preserves the time-series dependency of energy consumption and avoids the decline in model generalization ability caused by random shuffling.

[0192] 7. Model initialization: Based on the characteristics of enterprise production data, the parameter settings are as follows:

[0193] 7.1) Learning rate:

[0194] The initial learning rate was 0.05, balancing model training speed and overfit control. This avoided both excessively low learning rates leading to long training cycles and excessively high learning rates leading to model overfitting. After initialization, the model training cycle was controlled within 24 hours, and the convergence effect was good.

[0195] 7.2) Number of Decision Trees and Number of Leaves:

[0196] The number of decision trees (n_estimators) was initialized to 800, and the number of leaves (num_leaves) was set to the default value of 31. This ensures the model's fitting ability and avoids overfitting caused by an overly complex tree structure. After initialization, the model's fitting accuracy initially reached R²=0.88.

[0197] 7.3) Minimum data size in leaf nodes (min_data_in_leaf):

[0198] (1) Key production equipment (bending machine, refrigerant filling machine, leak detection equipment): min_data_in_leaf=40, in the range of 30~50, to avoid fitting operation noise.

[0199] (2) Auxiliary equipment (air compressor, cooling tower): min_data_in_leaf=15, which is between 10 and 20, taking into account both fitting accuracy and noise control.

[0200] 7.4) Sampling rate settings: The sample sampling rate (subsample) and feature sampling rate (colsample_bytree) are both set to 0.8 to reduce the risk of model overfitting. After initialization, the difference in R² between the training set and the validation set is controlled within 0.05.

[0201] 8. Feature Model Training and Parameter Tuning:

[0202] 8.1) Use the training set for model fitting and the validation set for real-time monitoring of model performance: Stop model training when the validation set RMSE does not decrease for 50 consecutive rounds to avoid overfitting. In actual training, the model converged in the 243rd round, and the validation set RMSE stabilized at 1.28kWh, at which point training was stopped, improving training efficiency by 35%.

[0203] 8.2) Optimize parameter tuning using a single-stage Bayesian optimization method combined with temporal cross-validation. Set the number of parameter trials n_trials=150, first coarsely search the parameter range, then finely search for the optimal solution, with the goal of minimizing the validation set RMSE.

[0204] (1) Coarse search stage: set learning_rate to 0.01~0.1, n_estimators to 600~1000, max_depth to 3~10, and filter out the optimal parameter range: learning_rate to 0.04~0.06, n_estimators to 700~900, max_depth to 5~7.

[0205] (2) Fine search stage: The search is refined within the coarse search interval to obtain the optimal parameter combination. The RMSE of the validation set is reduced from 1.28kWh to 0.95kWh, and the optimization effect is significant.

[0206] 8.3) Key parameter optimization: Based on temporal cross-validation, n_estimators, max_depth, and min_child_samples are used to improve the model's generalization ability.

[0207] (1) n_estimators=850, max_depth=6, key equipment min_child_samples=45, auxiliary equipment min_child_samples=18.

[0208] (2) After optimization, the R² of the model on the validation set increased from 0.88 to 0.92.

[0209] 8.4) Incorporate the historical energy consumption lag order k and the rolling window size into the hyperparameter search space for optimization:

[0210] (1) Short-term forecast (15min): The optimal lag order k=4, and the optimal size of the rolling window is 1h, 4h, 24h.

[0211] (2) Medium-term forecast (1d): The optimal lag order k=18, and the optimal size of the rolling window is 24h, 48h, and 72h.

[0212] After optimization, the model's prediction accuracy, especially its short-term real-time prediction accuracy, was improved, and the RMSE was reduced to 0.89kWh.

[0213] 9. A comprehensive evaluation of the model training results was conducted using the test set data, and the results are as follows:

[0214] 9.1) The test set data evaluation results are: RMSE=0.91kWh, MAE=0.73kWh, R²=0.92, WMAPE=0.85%, which meets the ideal target of R²≥0.90, and the model performance meets the standard.

[0215] 9.2) The contribution of features is quantified by SHAP value, and the importance of features is ranked as follows: output (SHAP value 0.32) > workshop temperature (SHAP value 0.25) > equipment uptime (SHAP value 0.18) > peak and off-peak electricity price (SHAP value 0.12) > number of process changeovers (SHAP value 0.09).

[0216] 9.3) Independent validation was performed using the test set (1728 data points), and an interpretable evaluation report was output, including model parameters, feature importance, prediction accuracy, error analysis, optimization suggestions, etc., to provide data support for energy management decisions.

[0217] 10. Based on the model evaluation results, identify high-energy-consuming equipment and formulate targeted energy-saving strategies:

[0218] 10.1) Determine the degree of association impact:

[0219] (1) Production fluctuations are the primary factor causing changes in energy consumption. For every 10 units / hour increase in production, electricity consumption increases by approximately 8.5 kWh.

[0220] (2) For every 1°C increase in workshop temperature, the hourly power consumption of the workshop increases by 2.6%, which is equivalent to an absolute value of about 3.2 kWh / °C. In summer, due to the higher workshop temperature, the average energy consumption is about 12.3% higher than that in spring and autumn.

[0221] (3) For every 1 hour increase in equipment operating time, energy consumption increases by approximately 4.8%; for every 0.1 decrease in equipment COP value, energy consumption increases by 2.1%.

[0222] (4) Unreasonable energy consumption during peak and off-peak electricity pricing periods and frequent process switching are also important reasons for high energy consumption.

[0223] 10.2) Plot the time series comparison curve of the test set prediction and actual value. The error between the predicted value and the actual value is mostly controlled within ±1kWh. The abnormal warning threshold can be set to ±10% of the predicted value. For example, if the predicted power consumption of the final assembly workshop is 130kWh / 15min, an abnormal alarm will be triggered when the actual power consumption is ≥143kWh or ≤117kWh.

[0224] Through comparative analysis, the No. 3 air compressor unit was identified as a high-energy-consuming device with a COP of 4.1, which is 8.7% higher than other air compressor units.

[0225] 10.3) Based on energy consumption attribution analysis, develop targeted energy-saving strategies:

[0226] (1) Peak-valley electricity price dispatch: Adjust the start-up time of key equipment and prioritize the start-up time of high energy-consuming equipment during off-peak hours (23:00~7:00).

[0227] (2) Optimize the start-up and shutdown times of key equipment according to production demand to avoid equipment idling.

[0228] (3) In winter, the workshop temperature is controlled at 22~24℃ by utilizing the waste heat from production.

[0229] (4) Optimize production scheduling to match output with equipment load and reduce the number of process switching.

[0230] 11. Integrate with other production-related information systems within the company to complete model deployment and system integration:

[0231] (1) Offline deployment: Daily energy consumption prediction is executed at 23:00 every day. The model is stored in encrypted Joblib format and equipped with hash verification to prevent tampering.

[0232] (2) Online deployment: Real-time inference service is encapsulated based on FastAPI, with identity authentication enabled on the interface and push delay ≤30s, for real-time scheduling in the workshop.

[0233] (3) Connect to the EMS energy management system to push energy consumption prediction data and abnormal alarm information, and realize the linkage between energy consumption prediction and energy monitoring. Connect to the MES production execution system to obtain real-time production data, adjust the parameters of the energy consumption prediction model, and realize the collaborative optimization of production and energy consumption. Connect to the APS production scheduling system to push energy consumption prediction data, support production scheduling optimization, and match production scheduling with energy consumption budget.

[0234] (4) Push abnormal alarms to WeChat for enterprise, with a response time of ≤5min, and support operation log traceability.

[0235] 12. Establish full lifecycle management for models to ensure their monitoring, iteration, and long-term usability:

[0236] 12.1) Model Application:

[0237] (1) Energy consumption anomaly warning: When the actual energy consumption deviates from the predicted value by ±10%, an anomaly alarm is triggered. For example, on a certain day, the actual power consumption of production line No. 1 in the sheet metal workshop was 145kWh / 15min, while the predicted value was 130kWh / 15min, which deviated by 11.5%, triggering an alarm. After investigation, it was found to be an equipment failure. After timely handling, further energy waste was avoided, which is consistent with the anomaly warning application scenario of similar enterprises.

[0238] (2) Production scheduling and energy consumption optimization: Combine energy consumption forecast data to optimize production scheduling. For example, high energy consumption processes are scheduled during off-peak hours and low energy consumption processes are scheduled during peak hours. The production schedule is optimized 3 times a month, and it is estimated that energy consumption will be saved by 28,000 kWh per month.

[0239] 12.2) Real-time monitoring: When the RMSE of the test set increases by ≥15% relative to the benchmark, or the R² decreases by ≥5% relative to the benchmark for 3 consecutive days, a precision drift alarm is triggered.

[0240] 12.3) Model update triggering conditions:

[0241] (1) Accuracy decline trigger: If the relative decline in prediction performance compared with the past 90 days exceeds 20% and lasts for 14 days, model retraining is triggered. In September, R² dropped from 0.93 to 0.73, a decrease of 21.5%, and lasted for 14 days, triggering the retraining condition.

[0242] (2) Triggering of changes in working conditions: When there are major changes in working conditions such as production line upgrades, equipment replacements, and process changes, retraining is triggered to ensure that the model can adapt to the new production conditions.

[0243] 12.4) Model Iteration and Version Management:

[0244] (1) Iteration frequency: monthly incremental fine-tuning (fine-tuning model parameters based on the newly added data in the current month, with a time consumption of ≤8h), and quarterly full retraining (retraining the model based on the data of the past 90 days, with a time consumption of ≤24h).

[0245] (2) Version management: Establish a model version management mechanism to record the parameters, performance and iteration time of each iteration, and support version rollback. For example, when the accuracy of the new iteration version decreases, it can be rolled back to the previous version.

[0246] (3) Accuracy assurance: Through iterative optimization, the model's mean absolute percentage error (MAPE) is guaranteed to be ≤10%. In actual operation, the MAPE is stable at 2.95%~8.75%, which meets production requirements.

[0247] Although the present invention has been described herein with reference to illustrative embodiments, the above embodiments are merely preferred embodiments of the present invention, and the implementation of the present invention is not limited to the above embodiments. It should be understood that those skilled in the art can devise many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A method for analyzing and predicting energy consumption data of an air conditioning manufacturing enterprise, characterized in that, Includes the following steps: Determine the indicators, scope, and cycle for collecting, analyzing, and predicting energy consumption data in air conditioning production; Define the collection boundaries for five types of structured and unstructured energy consumption data to avoid data overlap between different dimensions; Accessing and cleaning energy consumption data at the edge to prevent data distortion, timing misalignment, and intrusion of dirty data; Feature extraction of energy consumption data; Based on the temporal fluctuations and multi-factor coupling characteristics of energy consumption in air conditioning production, an energy consumption analysis and prediction feature model is constructed. Partitioning and cross-validating the time-series dataset; Based on the energy consumption differences of the main and auxiliary equipment in the air conditioning production line, the initial configuration of the model was completed, and a balance was sought between fitting power and noise-resistant generalization. The optimization objective is to minimize the root mean square error (RMSE) of energy consumption prediction on the validation set. The model is iteratively verified and hyperparameters are tuned. Quantitative metrics are used to evaluate the model training results; The independent test set that was not involved in training and parameter tuning was isolated throughout the process to complete the generalization ability test of the model, form a feature attribution ledger and anomaly error ledger, and provide an interpretable energy consumption model evaluation report. By combining real working conditions, energy consumption can be traced through model prediction, and energy-saving optimization and management strategies that can be implemented at the team and workshop levels can be output. The model was deployed and put into operation by using edge-based local deployment and integration with multiple business information systems.

2. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, The determination of the indicators, scope, and cycle for collecting, analyzing, and predicting energy consumption data in air conditioning production specifically includes: Quantitative monitoring indicators: covering energy consumption accounting indicators of the core air conditioning production workshop, including electricity / water / gas consumption per unit area and comprehensive energy consumption per unit product, eliminating invalid production capacity interference from reworked and repaired non-standard products and test prototypes; energy efficiency indicators of key power equipment, including the full-condition energy efficiency ratio of bending forming machines, fully automatic refrigerant charging machines, whole-machine leak detection equipment, and air compressor units, as well as the energy efficiency indicators of supporting equipment such as chiller units and circulating water pump rooms, to achieve traceability of energy consumption of main and auxiliary equipment; Monitoring scope: Achieve four levels of energy consumption monitoring scope: "total company - workshop - production line - key equipment", and link production process points, energy metering points and environmental sensing points to adapt to complex working conditions of staggered production in multiple workshops; The operating cycles are tiered: short-term sampling frequencies are 15min / 1h / 4h to meet the dynamic energy consumption monitoring needs of instantaneous equipment cluster start-up and shutdown and process switching in the workshop; medium-term sampling frequencies are 1d / 7d for shift scheduling and monthly rigid energy consumption quotas; long-term sampling frequencies are 30d / 90d for the technical transformation of high-energy-consuming equipment and medium- and long-term energy-saving and carbon reduction planning.

3. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, The determination of the collection boundaries for five types of structured and unstructured energy consumption data, to avoid data overlap across different dimensions, specifically includes: Basic energy consumption data: real-time / cumulative data from water, electricity, and gas meters. Data time sequence alignment, initial screening, and archiving are completed locally at the edge to prevent data latency and packet loss in the cloud. Production operation and planning data: ERP planned output, MES system actual qualified output, standard process rated production cycle time, equipment comprehensive efficiency (OEE) hourly statistics and running time, shift rotation log, start and stop of upstream and downstream process switching, production change shutdown time, annual overhaul shutdown period, and simultaneously mark abnormal working conditions such as temporary shutdown and emergency failure shutdown. Environmental data: including workshop temperature and humidity, outdoor temperature and humidity, and atmospheric pressure, used to adapt to the strong correlation between environmental temperature and humidity and the energy consumption of coupled equipment; Equipment operation and maintenance data: including speed, power, steam pressure, cooling water flow rate, and maintenance fault records; Energy-related data: peak-valley electricity prices and energy consumption allocation for projects.

4. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, The access and cleaning of edge-side energy consumption data, to prevent data distortion, timing misalignment, and intrusion of dirty data, includes the following specific methods: Deploy edge computing power, adopt 5G IIoT acquisition terminals, reuse on-site edge computing gateways, and complete the parsing, time-series normalization, and noise reduction preprocessing of raw energy consumption data at the nearest workshop location. Only transmit structured and effective data to the cloud data acquisition platform to reduce transmission pressure. By connecting to the EMS energy management system, MES production execution system, and SCADA equipment monitoring platform through MQTT and OPC UA protocols, the system can achieve access to multi-source heterogeneous data and ledger archiving. Hierarchical missing value handling: For instantaneous data loss caused by random instrument offline, interpolation of nearby points is used to fill in the missing data; for continuous missing data exceeding 3 sampling periods, the missing data is filled in by linking the equipment start / stop log and tracing the process production status; for planned shutdown conditions of equipment with complete power failure, the corresponding time series field is set to "0" and archived; for non-production conditions such as equipment power-on standby and low-load hibernation, the average steady-state standby energy consumption of the same shift, the same period, and the same equipment in the past 7 days is retrieved for backfilling; for periods marked with production abnormalities, emergency faults, etc., they are directly removed. Outlier handling: The "3σ+adaptive box plot" quartile check with a time-series sliding window is adopted instead of a single screening mechanism to adapt to the non-stationary, intermittent pulse-like fluctuation time-series characteristics of energy consumption in air conditioning production, and to eliminate outliers caused by sensor drift and sudden transmission interruptions, but should retain energy consumption pulse fluctuations caused by normal equipment start-up and shutdown and concentrated production. Redundant data processing: Based on MD5 global hash encoding, the uniqueness of timestamps and device location codes is verified synchronously to ensure the validity of the modeling dataset in a single time series, single device, and single entry, and to avoid interference from duplicate and redundant data.

5. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, The feature extraction of energy consumption data specifically includes: Time feature regularization: Unify the background sampling time benchmark, and extract hourly tags, weekly production schedule tags, monthly production start tags, statutory holiday shutdown tags, and work group shift variable features hourly; Nonlinear key feature transformation: Natural logarithmic transformation is applied to nonlinear features; sine-cosine orthogonal periodic encoding is used for strongly periodic time series features to conform to the energy consumption rhythm fluctuation law; For feature data that already belongs to the [0,1] interval, it is directly incorporated into the model; while for other cross-level and multidimensional related features, the "Min-Max" algorithm is uniformly used to compress and map them to [0,1].

6. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, The energy consumption analysis and prediction feature model is constructed based on the time-series fluctuations and multi-factor coupling characteristics of air conditioning production energy consumption, specifically including: Multi-source fusion benchmark features: integrate core production conditions, environmental features, time-series statistical features and calendar time-series features; process-energy consumption strong correlation interaction features to calculate the correlation coefficient between production cycle and instantaneous energy consumption, the coupling coefficient between equipment start-up time and comprehensive operating energy efficiency, the linkage strength between switching between upstream and downstream processes and instantaneous fluctuations in energy consumption, and superimposed peak-valley time-sharing linkage features. Time-series historical lag backtracking features: Construct a time-series energy consumption lag feature sequence, and adaptively determine the optimal gradient energy consumption lag order k based on the actual sampling frequency and prediction duration constraints; For short-term prediction of 15 minutes under typical working conditions, match the inertial fluctuation pattern of instantaneous energy consumption in the workshop. Scrolling window steady-state statistical characteristics: A fixed statistical window is used to calculate the extreme values, mean, standard deviation, and time-series slope of energy consumption within the window in real time, so as to draw a visual curve of energy consumption fluctuations and migration over time; Time-series differential dynamic characteristics: Calculate the energy consumption difference between the current moment and the previous moment hourly. It can intuitively represent instantaneous changes and pulse fluctuations, and quickly capture the energy consumption changes brought about by the start-up and shutdown of equipment clusters and centralized production.

7. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, The partitioning and cross-validation of the time-series dataset specifically includes: The samples were divided into a model training set, a hyperparameter optimization validation set, and an independent generalization test set according to the time interval. In terms of model validation, an incremental expansion of the segmentation window is adopted instead of random sampling validation to adapt to the real working conditions of rolling iteration on site. At the same time, a time-series cross-validation mechanism is used to preserve the time-series dependencies and conform to the energy consumption pattern of air conditioning production capacity during peak and off-peak seasons.

8. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, Based on the energy consumption differences between the main and auxiliary equipment in the air conditioning production line, the initial model configuration is completed, aiming to achieve a balance between fitting power and noise-resistant generalization. Specifically, this includes: Learning rate gradient: The initial learning rate baseline range is 0.01~0.1, taking into account both the model's gradient iteration convergence stability and computational limitations; Decision tree parameters: The preset number of decision trees n_estimators is in the baseline range of 600~1000, and the number of leaf nodes num_leaves is the default value of 31 for industrial modeling; Single leaf parameters: To adapt to different working conditions, the minimum sample size of leaves for core equipment is set to 28~50 to enhance the capture of fine energy consumption features; the minimum sample size of leaves for auxiliary equipment is set to 10~20 to effectively suppress random noise on site. To prevent model overfitting, the row dimension sample sampling rate (subsample) and column dimension feature sampling rate (colsample_bytree) are uniformly set to "0.8" in a single round of iterative training. And / or, with the goal of minimizing the root mean square error (RMSE) of energy consumption prediction on the validation set, iterative model verification and hyperparameter tuning are performed, specifically including: The training set is used to complete the model fitting iteration, the generalization error is monitored through the validation set, and the early stopping mechanism of early_stopping_rounds is enabled. Set the number of validation trials n_trials, first roughly screen out the reasonable range of parameters, and then finely search for the optimal hyperparameter combination to reduce the RMSE of the validation set; Gradually optimize the number of decision trees n_estimators, the maximum depth of a single tree max_depth, and the minimum number of leaf samples min_child_samples; The two core time-series features, the optimal energy consumption lag order k and the rolling statistical window size, are simultaneously incorporated into the hyperparameter search space to achieve mutual synergy between model structure and feature time series.

9. The method for analyzing and predicting energy consumption data of air conditioning manufacturing enterprises according to claim 1, characterized in that, The method combines real-world working conditions with model prediction to trace energy consumption sources and outputs energy-saving optimization and management strategies that can be implemented at the team and workshop levels. Specifically, these strategies include: By combining the importance ranking of the model's original features with the ranking of SHAP values, a process-energy consumption correlation ledger is established, and typical thresholds are identified. The system plots a time-series comparison curve of predicted and actual energy consumption values, identifies abnormal deviations from the warning threshold for energy consumption, and locates inefficient equipment and high-energy-consuming processes. It inputs monthly production capacity fluctuations and seasonal temperature change ranges, outputs the predicted energy consumption probability distribution, provides risk assessment, and completes dynamic load scheduling. Develop peak-valley staggered production scheduling plans and provide optimization suggestions for the staggered start-up and shutdown of high-energy-consuming equipment; implement zoned collaborative control of temperature and humidity for workshop process management.

10. A method for analyzing and predicting energy consumption data of an air conditioning manufacturing enterprise according to any one of claims 1-9, characterized in that, The method also includes: establishing a full lifecycle model monitoring system, gradient iterative optimization, and version backtracking to ensure the stability and long-term application of annual energy consumption forecasts.

Images