Energy consumption anomaly analysis method and system for environmental sanitation industry
By combining multivariate data collection with Large Language Modeling (LLM), the problem of high reliance on manual labor and difficulty in multi-factor coupling analysis in energy consumption management in the sanitation industry was solved, achieving efficient anomaly detection and accurate source tracing, and reducing the false alarm rate.
Patent Information
- Application Number
- CN202511159858.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-19
- Publication Date
- 2025-11-21
AI Technical Summary
Energy consumption management in the sanitation industry suffers from problems such as strong reliance on manual labor, difficulty in multi-factor coupling analysis, and weak traceability, resulting in low efficiency in anomaly detection and a high false alarm rate.
The method combines real-time acquisition of multivariate data, multimodal feature fusion, and large language model (LLM). A comprehensive feature vector F(t) is generated through a fusion function, and a pre-trained large language model (LLM) is used for anomaly diagnosis to generate a visual report to guide on-site verification.
It improves the accuracy of anomaly detection, accurately pinpoints the date of anomalies, reduces the false alarm rate, and outputs interpretable labels for equipment failures, false data, and sudden environmental changes to guide targeted handling.
Smart Images
Figure CN120997016A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of environmental sanitation energy consumption analysis technology, specifically involving a method and system for analyzing energy consumption anomalies in the environmental sanitation industry. Background Technology
[0002] Large AI models (or simply "large models") refer to a class of artificial intelligence models with a large number of parameters built from artificial neural networks. Large AI models are a relatively new concept that has emerged in the last decade. They are typically pre-trained on massive amounts of data using self-supervised or semi-supervised learning, and then their performance and capabilities are further optimized through fine-tuning based on instructions and human alignment. Large models are characterized by a large number of parameters, large amounts of training data, and large computational resources, and possess the ability to solve general tasks, follow human instructions, and perform complex reasoning.
[0003] The existing energy consumption management in the sanitation industry suffers from the following pain points: 1. High dependence on manual labor: Water, electricity and fuel consumption anomalies mainly rely on manual experience for judgment, which is inefficient and prone to missed detection; 2. Difficulty in multi-factor coupled analysis: Energy consumption is dynamically affected by multiple factors such as weather, season, traffic flow, and equipment quantity, and the traditional threshold method has a high false alarm rate; 3. Weak traceability: It is impossible to accurately locate the time point and root cause of the anomaly, making it difficult to support energy-saving optimization. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for analyzing energy consumption anomalies in the sanitation industry.
[0005] The technical solution adopted to solve the above technical problem includes the following specific steps:
[0006] Step 1: Real-time collection of multi-source data, collecting time-series energy consumption data of the target area, including water and electricity consumption and vehicle fuel consumption;
[0007] Step 2: Synchronize and correlate factor data, including pedestrian flow, number of devices in use, weather impact factors, and baseline data vectors for similar areas;
[0008] Step 3: Multimodal feature fusion, dynamically generating a comprehensive feature vector F(t) through a fusion function;
[0009] Step 4: Large model anomaly diagnosis. Input F(t) into the pre-trained large language model LLM, output the anomaly judgment result, identify the anomaly time point, generate anomaly type labels, and quantify the anomaly confidence.
[0010] Step 5: Generate a location report, outputting a visual report containing the time, type, and related factors of the anomaly to guide on-site verification.
[0011] The above technical solution dynamically couples factors such as weather, pedestrian flow, and equipment through a multimodal feature fusion function to eliminate environmental interference and significantly improve the accuracy of anomaly detection. The large language model outputs actionable conclusions, accurately pinpoints the date of anomalies, and analyzes the contribution of related factors based on the attention mechanism to output three types of interpretable labels: equipment failure, false data, and environmental mutation, guiding targeted handling. Dynamic time warping (DTW) adaptively matches similar regional benchmarks to reduce the false alarm rate and adapt to the needs of sanitation stations of different sizes.
[0012] Furthermore, the multimodal feature fusion adopts the following specific formula:
[0013] F(t) = α.STL(U t )+β.G(W t [C] p W e H f ])+γ.DTW(M t ,S t )
[0014] Among them, STL() is a seasonal time series decomposition algorithm that extracts the trend, periodicity, and residuals of hydropower and fuel consumption data; G() is an environmental factor weighted network based on an attention mechanism; DTW() is a dynamic time warping algorithm; and U t C represents the original monitoring values for water, electricity, and fuel consumption at all times. p For human traffic, W e H is a weather influencing factor. f S represents the number of devices in use. t M is the reference data vector for similar regions. t This is a vector of historical data for the regional baseline.
[0015] The above technical solutions can be used to fuse and analyze multimodal features, thereby improving monitoring accuracy.
[0016] Furthermore, the W e ∈[0,1], where 0.9 is for heavy rain and 0.1 is for sunny days, and α,β,γ are adjustable weighting coefficients.
[0017] The above technical solutions can be adjusted for different environments to avoid false alarms.
[0018] Furthermore, the G() network implements attention weighting in the following way:
[0019] Construct the fully connected layer to calculate the environmental factor weight matrix:
[0020]
[0021] Where Q = W e K = [Cp H f V is the feature vector, and then the weighted environmental feature vector is output.
[0022] The above technical solution allows for attention weighting of all environmental factors.
[0023] Furthermore, the large language model LLM is fine-tuned with data from the sanitation field, including the injection of sanitation equipment parameter library, energy consumption history case library, and fault knowledge graph, and the abnormal pattern recognition capability is optimized by instruction fine-tuning.
[0024] The above technical solutions can make large language models more adaptable to sanitation equipment.
[0025] A system for analyzing energy consumption anomalies in the sanitation industry includes a data perception layer, a feature fusion engine, a large model analysis layer, and a visualization interaction layer. The data perception layer includes smart water and electricity meters, vehicle OBD terminals, and a multimodal sensor network to collect real-time data on water and electricity consumption, vehicle fuel consumption and mileage, pedestrian flow, equipment operating status, and temperature and humidity.
[0026] The above technical solutions can collect comprehensive and accurate data, providing accurate basis for subsequent analysis.
[0027] Furthermore, the feature fusion engine includes an STL temporal decomposition module, an attention-weighted network module, and a DTW alignment module. The STL temporal decomposition module is responsible for parsing the trend, period, and residual components of energy consumption data. The attention-weighted network module is responsible for calculating the dynamic weights of environmental factors. The DTW alignment module is responsible for matching the current data with the benchmark pattern of similar regions.
[0028] The above technical solution can be used to weight and fuse the collected data, thereby enabling data processing.
[0029] Furthermore, the large model analysis layer includes a pre-trained LLM core, which loads fine-tuned parameters for the sanitation field; an anomaly detection submodule, which outputs anomaly time identifiers and type labels; a confidence calculation submodule, which generates probabilistic evaluation results; and a visualization interaction layer responsible for generating a multi-dimensional data comparison dashboard and a locatable report.
[0030] The above technical solutions can generate visual and comprehensive reports, making it easier for staff to process them.
[0031] The beneficial effects of this invention are as follows: This invention dynamically couples factors such as meteorology, human flow, and equipment through a multimodal feature fusion function to eliminate environmental interference, significantly improving the accuracy of anomaly detection. The large language model outputs actionable conclusions, accurately pinpointing the date of anomalies. Based on the attention mechanism, it analyzes the contribution of related factors and outputs three types of interpretable labels: equipment failure, false data, and environmental mutation, guiding targeted handling. Dynamic time warping (DTW) adaptively matches similar regional benchmarks to reduce the false alarm rate, adapting to the needs of sanitation stations of different sizes. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the hydropower anomaly analysis process of the present invention;
[0033] Figure 2 This is a schematic diagram of the fuel consumption anomaly analysis process of the present invention. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0035] like Figure 1 and Figure 2 As shown in this embodiment, a method and system for analyzing energy consumption anomalies in the sanitation industry includes the following specific steps:
[0036] Step 1: Real-time collection of multi-source data, collecting time-series energy consumption data of the target area, including water and electricity consumption and vehicle fuel consumption;
[0037] Step 2: Synchronize and correlate factor data, including pedestrian flow, number of devices in use, weather impact factors, and baseline data vectors for similar areas;
[0038] Step 3: Multimodal feature fusion, dynamically generating a comprehensive feature vector F(t) through a fusion function;
[0039] Step 4: Large model anomaly diagnosis. Input F(t) into the pre-trained large language model LLM, output the anomaly judgment result, identify the anomaly time point, generate anomaly type labels, and quantify the anomaly confidence.
[0040] Step 5: Generate a location report, outputting a visualized report containing the time, type, and related factors of the anomaly to guide on-site verification. By dynamically coupling factors such as weather, pedestrian flow, and equipment through a multimodal feature fusion function, environmental interference is eliminated, significantly improving the accuracy of anomaly detection. The large language model outputs actionable conclusions, accurately pinpointing the date of the anomaly. Based on the attention mechanism, the contribution of related factors is analyzed, and three types of interpretable labels are output: equipment failure, data false alarm, and environmental mutation, guiding targeted handling. Dynamic time warping (DTW) adaptively matches similar area benchmarks to reduce the false alarm rate, adapting to the needs of sanitation stations of different sizes.
[0041] The multimodal feature fusion adopts the following specific formula:
[0042] F(t) = α.STL(U t )+β.G(W t [C] p W e H f ])+γ.DTW(M t ,S t )
[0043] Among them, STL() is a seasonal time series decomposition algorithm that extracts the trend, periodicity, and residuals of hydropower and fuel consumption data; G() is an environmental factor weighted network based on an attention mechanism; DTW() is a dynamic time warping algorithm; and U t C represents the original monitoring values for water, electricity, and fuel consumption at all times. p For human traffic, W e H is a weather influencing factor. f S represents the number of devices in use. t M is the reference data vector for similar regions. t Using regional baseline historical data vectors, multimodal features can be fused and analyzed to improve monitoring accuracy.
[0044] The W e ∈[0,1], where 0.9 is for heavy rain and 0.1 is for sunny days. α, β, γ are adjustable weight coefficients that can be adjusted for different environments to avoid false alarms.
[0045] The G() network implements attention weighting in the following way:
[0046] Construct the fully connected layer to calculate the environmental factor weight matrix:
[0047]
[0048] Where Q = W e K = [C p H fV is the feature vector, and then the weighted environmental feature vector is output, which can perform corresponding attention weighting for all environmental factors.
[0049] The large language model LLM is fine-tuned with data from the sanitation field, including the injection of sanitation equipment parameter library, energy consumption history case library, and fault knowledge graph. The abnormal pattern recognition capability is optimized by using instruction fine-tuning, which can make the large language model more suitable for sanitation equipment.
[0050] A system for analyzing energy consumption anomalies in the sanitation industry includes a data perception layer, a feature fusion engine, a large model analysis layer, and a visualization interaction layer. The data perception layer includes smart water and electricity meters, vehicle OBD terminals, and a multimodal sensor network, which collects water and electricity consumption data, vehicle fuel consumption and mileage data, pedestrian flow, equipment operating status, temperature and humidity data in real time. It can collect comprehensive and accurate data, providing accurate basis for subsequent analysis.
[0051] The feature fusion engine includes an STL temporal decomposition module, an attention-weighted network module, and a DTW alignment module. The STL temporal decomposition module is responsible for parsing the trend, period, and residual components of energy consumption data. The attention-weighted network module is responsible for calculating the dynamic weights of environmental factors. The DTW alignment module is responsible for matching the current data with the benchmark pattern of similar regions. It can perform weighted fusion of the collected data to achieve data processing.
[0052] The large model analysis layer includes a pre-trained LLM core, which loads fine-tuned parameters for the sanitation field; an anomaly detection submodule, which outputs anomaly time identifiers and type labels; a confidence calculation submodule, which generates probabilistic evaluation results; and a visualization interaction layer, which is responsible for generating multi-dimensional data comparison dashboards and location-based reports, enabling the generation of visual and comprehensive reports to facilitate corresponding processing by staff.
[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.
Claims
1. A method for analyzing energy consumption anomalies in the sanitation industry, characterized in that, The specific steps include the following: Step 1: Real-time collection of multi-source data, collecting time-series energy consumption data of the target area, including water and electricity consumption and vehicle fuel consumption; Step 2: Synchronize and correlate factor data, including pedestrian flow, number of devices in use, weather impact factors, and baseline data vectors for similar areas; Step 3: Multimodal feature fusion, dynamically generating a comprehensive feature vector F(t) through a fusion function; Step 4: Large model anomaly diagnosis. Input F(t) into the pre-trained large language model LLM, output the anomaly judgment result, identify the anomaly time point, generate anomaly type labels, and quantify the anomaly confidence. Step 5: Generate a location report, outputting a visual report containing the time, type, and related factors of the anomaly to guide on-site verification.
2. The method for analyzing energy consumption anomalies in the sanitation industry according to claim 1, characterized in that, The multimodal feature fusion adopts the following specific formula: F(t)=α.STL(U t )+β.G(W t .[C p ,W e ,H f ])+γ.DTW(M t ,S t ) Among them, STL() is a seasonal time series decomposition algorithm that extracts the trend, periodicity, and residuals of hydropower and fuel consumption data; G() is an environmental factor weighted network based on an attention mechanism; DTW() is a dynamic time warping algorithm; and U t C represents the original monitoring values for water, electricity, and fuel consumption at all times. p For human traffic, W e H is a weather influencing factor. f S represents the number of devices in use. t M is the reference data vector for similar regions. t This is a vector of historical data for the regional baseline.
3. The method for analyzing energy consumption anomalies in the sanitation industry according to claim 2, characterized in that, The W e ∈[0,1], where 0.9 is for heavy rain and 0.1 is for sunny days, and α,β,γ are adjustable weighting coefficients.
4. The energy consumption anomaly analysis method 3 for the sanitation industry according to claim 3, characterized in that, The G() network implements attention weighting in the following way: Construct the fully connected layer to calculate the environmental factor weight matrix: Where Q = W e K = [C p H f V is the feature vector, and then the weighted environmental feature vector is output.
5. The method for analyzing energy consumption anomalies in the sanitation industry according to claim 4, characterized in that, The large language model (LLM) is fine-tuned with data from the sanitation field, including the injection of sanitation equipment parameter library, energy consumption history case library, and fault knowledge graph, and the abnormal pattern recognition capability is optimized by using instruction fine-tuning.
6. The system for energy consumption anomaly analysis in the sanitation industry according to claim 5, characterized in that, It includes a data perception layer, a feature fusion engine, a large model analysis layer, and a visualization interaction layer. The data perception layer includes smart water and electricity meters, vehicle OBD terminals, and a multimodal sensor network to collect real-time data on water and electricity consumption, vehicle fuel consumption and mileage, pedestrian flow, equipment operating status, and temperature and humidity.
7. The system for energy consumption anomaly analysis in the sanitation industry according to claim 6, characterized in that, The feature fusion engine includes an STL time series decomposition module, an attention-weighted network module, and a DTW alignment module. The STL time series decomposition module is responsible for parsing the trend, period, and residual components of energy consumption data. The attention-weighted network module is responsible for calculating the dynamic weights of environmental factors. The DTW alignment module is responsible for matching the current data with the benchmark pattern of similar regions.
8. The system for energy consumption anomaly analysis in the sanitation industry according to claim 7, characterized in that, The large model analysis layer includes a pre-trained LLM core, which loads fine-tuned parameters for the sanitation field. The anomaly detection submodule outputs the anomaly time identifier and type label; The confidence calculation submodule generates probabilistic evaluation results; the visualization interaction layer is responsible for generating a multi-dimensional data comparison dashboard and a location-based report.