An AI-based energy consumption data analysis and prediction system

Abstract

This invention relates to the field of energy consumption analysis and discloses an AI-based energy consumption data analysis and prediction system. The system includes collecting environmental parameters and operating status of the equipment, dynamically identifying the current system operating condition using an algorithm combining incremental clustering and historical pattern matching, extracting multi-scale features from the collected data according to time, space, and operating condition dimensions, and dynamically adjusting the normalization parameters as the operating condition changes. A drift index is calculated by comparing a drift threshold with historical distributions. Based on the drift index, an optimal normalization update strategy is selected to dynamically update the normalization parameters. A statistical model is used to predict energy consumption trends. The prediction results are combined with operating condition labels for weighted correction. Error analysis and deviation detection are performed by combining the drift index, operating condition prediction, and historical error data. The analysis results are fed back to the operating condition perception module, the feature adaptation module, and the normalization control module. This invention has the advantage of improving stability and reliability under dynamic environments.

Description

Technical Field

[0001] This invention relates to the field of energy consumption analysis, specifically to an AI-based energy consumption data analysis and prediction system. Background Technology

[0002] With the development of energy management technology, energy consumption data-based analysis and prediction systems have become important tools for building energy conservation, industrial energy efficiency optimization, and intelligent scheduling. These systems collect multi-dimensional energy consumption data such as temperature, humidity, load, current, and voltage to model and predict energy consumption patterns, thus assisting in energy optimization and operational decision-making. In existing technologies, to improve the stability and convergence speed of prediction model training, input features are typically normalized or standardized, such as through min-max normalization or mean-variance normalization. Normalization parameters are generally calculated based on historical data and kept fixed. However, the system operating environment and equipment conditions change over time, such as production line adjustments, seasonal changes, equipment aging, or load fluctuations, leading to shifts in the distribution of input features and making the original normalization parameters unable to accurately reflect the current data characteristics. When the normalization parameters do not match the current data distribution, feature scale imbalance occurs, and the feature space drifts, thus affecting prediction accuracy. This problem is particularly pronounced in long-term operation or multi-condition scenarios, and existing methods lack adaptive adjustment mechanisms for changes in input distribution. Therefore, it is necessary to design an AI-based energy consumption data analysis and prediction system that can improve stability and reliability in dynamic environments. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an AI-based energy consumption data analysis and prediction system, which has the advantages of improving stability and reliability in dynamic environments and solves the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of improving stability and reliability under dynamic environments, this invention provides the following technical solution: an AI-based energy consumption data analysis and prediction system, comprising:

[0005] Operating condition perception module: Collects environmental parameters and operating status of equipment, uses an operating condition recognition algorithm that combines incremental clustering and historical pattern matching to dynamically identify the current system operating condition, and outputs operating condition labels and future operating condition prediction results to the feature adaptive module.

[0006] Feature Adaptive Module: Receives working condition labels and future working condition prediction results, divides the collected data into time, space and working condition dimensions to extract multi-scale features, and performs weighted normalization processing based on working condition labels and prediction results. The normalization parameters can be dynamically adjusted as the working conditions change, and the processed features are output to the normalization control module.

[0007] Normalization control module: Receives features output from the feature adaptation module, monitors feature distribution drift, calculates drift index by comparing the drift threshold with historical distribution, selects the optimal normalization update strategy based on the drift index, dynamically updates the normalization parameters, and outputs the updated features to the prediction engine module;

[0008] Prediction Engine Module: Receives the output features from the normalized control module, uses a statistical model to predict energy consumption trends, combines the prediction results with operating condition labels for weighted correction, and outputs them to the feedback optimization module;

[0009] Feedback optimization module: Receives prediction results, combines drift index, operating condition prediction and historical error data to perform error analysis and deviation detection, and feeds back the analysis results to the operating condition perception module, feature adaptation module and normalization control module.

[0010] Preferably, the process of dynamically identifying the current system operating condition is as follows:

[0011] Collect real-time operating parameters and environmental data of the equipment, including temperature, humidity, load, current and voltage;

[0012] Incremental clustering and similarity matching are performed between the current data and historical operating conditions;

[0013] Output the current operating condition category and future operating condition trend prediction;

[0014] Monitor sudden changes in operating conditions, such as abnormal load fluctuations or environmental disturbances, and dynamically correct the operating condition labels. Transmit the identification results and prediction information to the feature adaptive module in real time.

[0015] Preferably, the process of extracting multi-scale features from the collected data according to time, space, and operating conditions is as follows:

[0016] The collected data is sliced ​​in multiple dimensions according to time window, equipment distribution and operating condition category;

[0017] Calculate statistical characteristics within each slice, including mean, variance, rate of change of peak value, and time series trend characteristics;

[0018] Statistical and trend features are calculated within each time-space slice and weighted and summarized according to working conditions to form a feature matrix that can be directly input into the prediction model.

[0019] Features under different working conditions are categorized and labeled, and the feature vectors of each scale are output to the feature adaptive module for weighted normalization.

[0020] Preferably, the weighted normalization process based on the working condition labels and prediction results is as follows:

[0021] Each multi-scale feature vector is assigned a weight coefficient related to the working condition label and future trend.

[0022] The features are scaled using a normalization function;

[0023] The normalization weights are dynamically adjusted based on the operating condition prediction results.

[0024] The normalization parameter changes in real time, and the weighted normalized features are output to the normalization control module.

[0025] Preferably, the normalized parameter can be dynamically adjusted according to changes in operating conditions as follows:

[0026] Real-time monitoring of the magnitude and drift trend of feature distribution changes;

[0027] Adjust the normalization parameters based on the current operating condition labels and prediction results, including the numerical settings and updates of the mean, variance, maximum and minimum values;

[0028] The normalized parameters are continuously updated using a sliding window or exponentially weighted moving average method.

[0029] The normalized parameter update process is logged and historically tracked to generate parameter evolution curves, and the updated normalized parameters are output to the normalization control module.

[0030] Preferably, the process of calculating the drift index by comparing the drift threshold with the historical distribution is as follows:

[0031] Perform statistical analysis on the weighted normalized eigenvectors to calculate the deviation between the current distribution and the historical distribution;

[0032] The drift index is generated by comparing it with the set drift threshold.

[0033] The impact of drift is assessed by combining feature importance and operating condition labels;

[0034] When the drift index exceeds the threshold, the normalization parameter update strategy selection is triggered, and the drift index and analysis results are transmitted to the normalization control module.

[0035] Preferably, the process of dynamically updating the normalized parameters is as follows:

[0036] Based on the feature drift magnitude and feature importance, either a sliding window update strategy or an exponentially weighted moving average strategy is selected to dynamically adjust the normalization parameters;

[0037] Adjust the mean, variance, and maximum / minimum parameters to perform weighted priority updates on key features;

[0038] The updated feature vector is output to the prediction engine module, and the history of normalization parameter adjustment is recorded at the same time.

[0039] Preferably, the process of predicting energy consumption trends using a statistical model is as follows:

[0040] Receive the feature vector output by the normalization control module;

[0041] Prediction models are selected based on historical energy consumption data and current characteristics, including linear regression, random forest, and deep neural networks;

[0042] The energy consumption trend is predicted based on the trained model, and the predicted value and confidence interval are output.

[0043] Monitor prediction deviations and abnormal fluctuations, and transmit the prediction results to the feedback optimization module.

[0044] Preferably, the weighted correction process combining the prediction results with the operating condition labels is as follows:

[0045] Based on the current operating condition labels and future operating condition trends, dynamic weighting coefficients are assigned to the prediction results;

[0046] A deviation analysis was conducted between the original prediction results and historical energy consumption data, combining prediction error statistics, drift indicators, and historical deviation distribution.

[0047] The correction results are optimized using weighted adjustment or minimum mean square error optimization methods.

[0048] The system records the weighting coefficients, correction magnitudes, and error changes at each step of the correction process in real time, generates correction logs and deviation evolution curves, and outputs the energy consumption prediction results after dynamic weighted correction and error optimization to the feedback optimization module.

[0049] Preferably, the process of feeding back the analysis results to the working condition sensing module, the feature adaptive module, and the normalization control module is as follows:

[0050] Statistical analysis of prediction errors is performed, and the deviation distribution is calculated by combining drift index and historical error.

[0051] Generate deviation reports and feed them back to the working condition perception module in real time to update working condition labels and prediction trends;

[0052] The deviation analysis results are transferred to the feature adaptive module to adjust the feature weighting coefficients and normalization parameters.

[0053] The drift index and deviation information are fed back to the normalization control module to update the normalization strategy and parameters.

[0054] Data feedback and parameter transfer are performed between modules to complete the normalization parameter update and feature weighting adjustment.

[0055] Compared with existing technologies, the present invention provides an AI-based energy consumption data analysis and prediction system, which has the following beneficial effects:

[0056] This invention achieves adaptive analysis and dynamic prediction of energy consumption data under complex operating conditions by constructing a multi-level intelligent analysis architecture that includes operating condition perception, feature adaptation, normalization control, prediction engine, and feedback optimization. The system can accurately perceive the operating status of equipment based on operating condition recognition algorithms, and through feature adaptation and normalization control mechanisms, dynamically normalizes and corrects drift for features under different time, space, and operating condition dimensions, thereby maintaining the stability of feature distribution. The prediction engine module combines statistical models and operating condition information to achieve accurate prediction of energy consumption trends, while the feedback optimization module completes closed-loop parameter updates and model self-tuning based on error analysis and drift indicators. Through the synergistic effect of each module, this system has the advantage of improving prediction stability and system reliability in dynamic environments, and can maintain the accuracy and adaptability of energy consumption prediction over a long period. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0058] 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.

[0059] Example 1: Please refer to Figure 1 As shown in the figure, an AI-based energy consumption data analysis and prediction system according to an embodiment of the present invention includes:

[0060] Operating condition perception module: Collects environmental parameters and operating status of equipment, uses an operating condition recognition algorithm that combines incremental clustering and historical pattern matching to dynamically identify the current system operating condition, and outputs operating condition labels and future operating condition prediction results to the feature adaptive module.

[0061] The process of dynamically identifying the current system operating condition in the operating condition perception module is as follows:

[0062] Collect real-time operating parameters and environmental data of the equipment, including temperature, humidity, load, current and voltage;

[0063] The system collects real-time operating parameters and environmental data, including temperature, humidity, load, current, and voltage, through sensor arrays deployed at key locations on the equipment. The temperature and humidity sensors have a sampling accuracy of 2%, and relative humidity changes can be recorded once per second, ensuring that high-frequency environmental changes are captured. The load sensors measure mechanical and electrical loads through current transformers and strain gauges, with a sampling rate of fifty times per second. Voltage and current measurements use high-precision acquisition modules to ensure that the numerical resolution reaches the millivolt and milliampere levels. All collected data is synchronized and stored through timestamps and equipment identifiers to form a unified data stream.

[0064] Incremental clustering and similarity matching are performed between the current data and historical operating conditions;

[0065] After standardization and denoising, the collected real-time data is input into an incremental clustering algorithm for operating condition classification. The incremental clustering algorithm uses a historical operating condition database as a reference to calculate the similarity between new data and existing operating condition patterns. It evaluates the degree of deviation of the current data from the center of each type of operating condition by using Euclidean distance or cosine similarity. During the calculation process, the system assigns weights to different features: temperature and humidity changes account for 20% of the total weight, load and current changes account for 50% of the weight, and voltage fluctuations account for 30% of the weight, to reflect the importance of different parameters to the operating condition determination. The algorithm adaptively updates for new operating conditions or abnormal patterns.

[0066] Output the current operating condition category and future operating condition trend prediction;

[0067] The incremental clustering results are used to generate current operating condition labels after thresholding. At the same time, trend prediction is performed by combining time series analysis and historical operating condition evolution patterns. The trend prediction adopts sliding time window analysis to calculate the linear regression slope and moving average of each parameter. For example, in the case of a sudden increase in equipment load, the system can predict that the load may increase by 15% to 20% in the next three minutes and convert the prediction results into trend labels. All operating condition labels and trend information are written to the data cache in chronological order and transmitted to the feature adaptive module in real time.

[0068] Monitor sudden changes in operating conditions, such as abnormal load fluctuations or environmental disturbances, and dynamically correct the operating condition labels. Transmit the identification results and prediction information to the feature adaptive module in real time.

[0069] The system monitors sudden operating conditions, including abnormal load fluctuations and environmental disturbances, by setting thresholds and anomaly detection algorithms. Specifically, it calculates the sliding standard deviation and peak rate of change of real-time acquired data. When the load change exceeds three times the historical average standard deviation or the temperature and humidity change exceeds 10%, an anomaly event is triggered. The anomaly event record includes the occurrence time, relevant parameter values, and deviation magnitude. The system also immediately triggers a dynamic correction algorithm for operating condition labels. The dynamic correction algorithm corrects the current operating condition label by recalculating the matching degree between the incremental cluster center point and the real-time data, so that the label can reflect the state changes caused by short-term disturbances in a timely manner. The identified operating condition labels and trend prediction information are sent to the feature adaptive module through a high-speed data bus.

[0070] Feature Adaptive Module: Receives working condition labels and future working condition prediction results, divides the collected data into time, space and working condition dimensions to extract multi-scale features, and performs weighted normalization processing based on working condition labels and prediction results. The normalization parameters can be dynamically adjusted as the working conditions change, and the processed features are output to the normalization control module.

[0071] The adaptive feature module divides the collected data into time, space, and operating condition dimensions to extract multi-scale features as follows:

[0072] The collected data is sliced ​​in multiple dimensions according to time window, equipment distribution and operating condition category;

[0073] The real-time data received by the system is first sliced ​​according to time. The length of each time window is set to one to five minutes depending on the dynamic characteristics of the equipment. At the same time, the start and end timestamps are recorded to ensure the continuity and traceability of the time series. In the spatial dimension, the sensor data from different physical locations are divided according to the equipment number and installation location, so that the signals from the same partition or unit within the same time window form spatial slices. In the operating condition dimension, the data of the same time and spatial slices are grouped according to the operating condition labels output by the operating condition sensing module, realizing a three-dimensional slice structure of time, space and operating condition. Each slice forms an independent data unit for subsequent feature extraction and statistical analysis.

[0074] Calculate statistical characteristics within each slice, including mean, variance, rate of change of peak value, and time series trend characteristics;

[0075] Within each 3D slice, the system first performs noise reduction and outlier processing on the raw data, using moving average filtering and median filtering to eliminate transient fluctuations. Subsequently, it calculates statistical features, including the mean, variance, and peak change rate, and extracts time-series trend features, such as short-term slope, moving average increase / decrease rate, and peak interval change. The mean reflects the overall level of the parameter, the variance reflects the fluctuation amplitude, the peak change rate reflects the sudden change trend, and the time-series trend features reveal the evolution of the parameter over time. These features are stored in matrix form, with each row representing the statistical features within a time window and each column representing the feature dimensions of different sensors or devices.

[0076] Statistical and trend features are calculated within each time-space slice and weighted and summarized according to working conditions to form a feature matrix that can be directly input into the prediction model.

[0077] The system extracts features from each time-space slice and weights them according to the current operating condition label and the future operating condition trend. The weight coefficients are determined by the historical importance of the operating condition label and the predicted value of the future trend. For example, the weight of load change increases under high load conditions, and the weight of temperature and humidity change increases under environmental disturbance conditions. The weighted summation is achieved through matrix multiplication and column vector weighting to obtain the comprehensive feature vector of each slice, ensuring that features of different dimensions reflect their importance in the overall input and forming a feature matrix that can be directly input into the prediction model.

[0078] The features under different working conditions are classified and labeled, and the feature vector of each scale is output to the feature adaptive module for weighted normalization.

[0079] The system categorizes and labels the weighted and aggregated feature vectors according to operating conditions. Each feature vector is accompanied by an operating condition label, a timestamp, and a spatial location identifier, ensuring that subsequent modules can identify the source and context of the features. The feature vectors are stored in a two-dimensional matrix, with each row representing a feature of a time-space slice and each column representing statistical or trend features of different dimensions. The operating condition label column indicates the operating condition category. The feature vectors with completed operating condition labeling are transmitted to the feature adaptation module via a high-speed data bus. The module performs multi-scale weighted normalization processing on the feature vectors, including mean adjustment, variance standardization, and amplitude scaling, to ensure that features of different scales and types maintain consistency in the input space of the prediction model. During transmission, the system records the timestamp, operating condition label, and spatial distribution information of each feature vector, ensuring that subsequent normalization parameter updates and drift monitoring can trace historical feature changes, forming a continuous data chain.

[0080] The weighted normalization process in the feature adaptive module, based on the working condition label and prediction results, is as follows:

[0081] Each multi-scale feature vector is assigned a weight coefficient related to the working condition label and future trend.

[0082] The system receives multi-scale feature vectors from the feature extraction module. Each vector contains statistical and trend features of time, space, and operating conditions. Based on the current operating condition label and future operating condition trend prediction results output by the operating condition perception module, a weight coefficient is calculated for each feature dimension. The weight coefficient is calculated by the contribution of the feature to the energy consumption prediction model under the operating condition in historical data. For example, the weight of load-related features is increased under high load conditions, and the weight of temperature and humidity-related features is adjusted under environmental disturbance conditions. The weight coefficients are stored in vector form and element-wise multiplied with the corresponding feature vectors to form a preliminary weighted feature matrix.

[0083] The features are scaled using a normalization function;

[0084] The weighted feature matrix enters the normalization process. The system calculates the mean and standard deviation for each column of features, and standardizes them by subtracting the mean and dividing by the standard deviation. At the same time, linear scaling is used for extreme features to ensure that the feature values ​​fall within the preset input range of the model. The normalization function can use a sliding window to calculate the mean and variance to reflect the dynamic changes of the features over time. For non-Gaussian distributed features, the system uses quantile normalization or min-max scaling to avoid outliers from causing deviations in the model input. The normalized feature matrix maintains the original feature dimension structure, with each row corresponding to a time-space slice and each column corresponding to a statistical or trend feature dimension.

[0085] The normalization weights are dynamically adjusted based on the operating condition prediction results.

[0086] The system combines the future operating condition trends output by the operating condition prediction module and dynamically adjusts the weights of the normalized features. For example, when it is predicted that the system will enter a high-power load condition in the near future, the weights of load-related features in the normalized matrix are increased, and the weights of temperature and humidity features are adjusted accordingly based on historical sensitivity. The dynamic weights are applied to the normalized feature matrix through multiplication coefficients or exponential weighting, so that the input features not only meet the standardization requirements but also reflect the forward-looking impact of operating condition prediction on energy consumption changes.

[0087] Record the changing trend of normalization parameters in real time and output the weighted normalized features to the normalization control module;

[0088] During the normalization process, the system records the changes in mean, standard deviation, minimum, maximum, and weighting coefficients of each feature in real time, forming time series data to track the dynamic evolution of the normalization parameters. The recorded data includes timestamps, operating condition labels, spatial locations, and feature dimension identifiers, and is stored in the parameter history database. After completing the weighted normalization process, the system transmits the processed feature matrix to the normalization control module via a high-speed data bus.

[0089] The process by which the normalized parameters in the adaptive feature module can be dynamically adjusted according to changes in operating conditions is as follows:

[0090] Real-time monitoring of the magnitude and drift trend of feature distribution changes;

[0091] After receiving the weighted and normalized feature matrix, the system performs real-time statistical analysis on each feature dimension, including mean, variance, maximum, minimum, distribution skewness, and kurtosis. It calculates the deviation between the current feature distribution and the historical feature distribution through a sliding time window, quantifies the drift amplitude, and calculates the drift trend through the rate of change of deviation over a continuous time period. This is used to identify the dynamic evolution of features as operating conditions change. The monitoring frequency is consistent with the data acquisition frequency, for example, updating statistical indicators once per second to ensure that short-cycle fluctuations in operating conditions can be captured. At the same time, historical statistical results are stored in the database in time series format to provide a reference for parameter updates.

[0092] Adjust the normalization parameters based on the current operating condition labels and prediction results, including the numerical settings and updates of the mean, variance, maximum and minimum values;

[0093] The system combines the current operating condition labels provided by the operating condition perception module with the future operating condition trends output by the prediction module to adjust the normalized parameters of each feature dimension. The adjustment method is to select the corresponding historical parameter template according to the operating condition label, compare the current real-time statistical indicators with the template, and update the mean, variance, and extreme value range. For situations where high load or temperature and humidity fluctuations are about to occur, the system adjusts the parameters according to historical statistical data so that the feature input can reflect the upcoming change in operating condition. The parameter adjustment is recorded in matrix form, with each column corresponding to a feature dimension and each row corresponding to the normalized parameters of the current time window.

[0094] The normalized parameters are continuously updated using a sliding window or exponentially weighted moving average method.

[0095] The system selects either a sliding window update or an exponentially weighted moving average method to continuously iterate the normalized parameters. The sliding window method calculates the mean and variance through a fixed-length historical window, incorporating the latest data into the window and discarding the earliest data. The exponentially weighted moving average method sets a decay coefficient for each feature dimension, making parameter updates more sensitive to recent changes by weighting historical parameters and current statistical indicators. The continuous update process ensures that the normalized parameters can dynamically respond to short-period disturbances and long-term drifts, while avoiding transient anomalies from causing biases in the model input.

[0096] Log the normalization parameter update process and track its history to form a parameter evolution curve. Output the updated normalization parameters to the normalization control module.

[0097] During parameter updates, the system records the mean, variance, maximum, minimum, and update timestamp of each feature dimension to the parameter log database. The log data is stored in time series format, forming a complete parameter evolution curve. This allows the system to track the dynamic trend of normalized parameters as the operating conditions change. The database also records operating condition labels, future prediction trends, and update method types, providing a data foundation for subsequent drift analysis, normalization strategy evaluation, and model input validation. After completing continuous updates and historical records, the system sends the latest normalized parameter matrix for each feature dimension to the normalization control module.

[0098] Normalization control module: Receives features output from the feature adaptation module, monitors feature distribution drift, calculates drift index by comparing the drift threshold with historical distribution, selects the optimal normalization update strategy based on the drift index, dynamically updates the normalization parameters, and outputs the updated features to the prediction engine module;

[0099] The process of calculating the drift index in the normalization control module by comparing the drift threshold with the historical distribution is as follows:

[0100] Perform statistical analysis on the weighted normalized eigenvectors to calculate the deviation between the current distribution and the historical distribution;

[0101] The system receives the weighted normalized feature vector output by the feature adaptive module and performs real-time statistical analysis on each feature dimension, including mean, variance, maximum, minimum, distribution skewness, and kurtosis. It compares the statistical indicators of the current time window with the statistical distribution under similar historical operating conditions. The drift amplitude of the current feature distribution is measured by calculating the mean difference, variance ratio, and extreme value deviation. The statistical calculation adopts the sliding time window method, and the window length can be dynamically set according to the equipment acquisition frequency and the rate of change of operating conditions to ensure that both short-term fluctuations and long-term trends are effectively captured. The deviation calculation results are stored in matrix form, with each row corresponding to a time window slice and each column corresponding to a feature dimension, providing basic data for the subsequent generation of drift indicators.

[0102] The drift index is generated by comparing it with the set drift threshold.

[0103] The system compares the calculated deviation of each feature dimension with a preset drift threshold one by one. The drift threshold is determined by historical operating data and system design parameters and can be set differently for different feature dimensions. During the comparison process, a binary drift identifier is generated for each feature dimension and further summarized into a comprehensive drift index matrix. The matrix contains the feature dimension identifier, deviation value and drift judgment status. The comprehensive drift index can be used to evaluate whether the overall feature distribution has drifted significantly and provide a quantitative basis for updating the normalization parameters.

[0104] The impact of drift is assessed by combining feature importance and operating condition labels;

[0105] The system weights the drift index of each feature dimension with the feature importance weight. The feature importance is determined by the training results of historical prediction models or expert evaluation. Combined with the current operating condition label, the system evaluates the potential impact of each feature drift on the overall energy consumption prediction accuracy. Through matrix operations, the drift amplitude, feature weight and operating condition label are integrated to generate a drift impact matrix, which provides a basis for dynamically selecting the normalization parameter update strategy and ensures that key feature drifts are given priority response.

[0106] When the drift index exceeds the threshold, the normalization parameter update strategy selection is triggered, and the drift index and analysis results are transmitted to the normalization control module.

[0107] The system performs threshold determination on the comprehensive drift index. When the index exceeds the set threshold, it automatically triggers the normalization parameter update strategy selection process. The update strategies include sliding window update and exponentially weighted moving average update. The appropriate method can be selected according to the drift amplitude and feature importance. The strategy selection process calculates the sensitivity of each method to feature distribution recovery through an algorithm and generates a parameter update scheme. The generated drift index matrix and detailed analysis results, including the deviation value of each feature dimension, drift determination status, feature weight and operating condition label information, are sent to the normalization control module through the data bus.

[0108] The process of dynamically updating the normalized parameters in the normalization control module is as follows:

[0109] Based on the feature drift magnitude and feature importance, either a sliding window update strategy or an exponentially weighted moving average strategy is selected to dynamically adjust the normalization parameters;

[0110] The system receives the drift index matrix calculated by the normalization control module, quantifies the drift amplitude of each feature dimension, and generates a comprehensive evaluation score by combining the feature importance weights. Based on the evaluation results, the system selects an appropriate normalization parameter update strategy. For dimensions with small drift amplitudes and low feature weights, a sliding window update strategy is adopted, which calculates the mean and variance through a fixed-length time window to achieve smooth updates. For key features with large drift amplitudes or high weights, an exponentially weighted moving average strategy is adopted, which assigns higher weights to the most recently collected data to improve the response speed to sudden changes in operating conditions. The strategy selection process is completed through matrix operations, and a strategy identifier table corresponding to each feature is generated to provide a decision basis for parameter updates.

[0111] Adjust the mean, variance, and maximum / minimum parameters to perform weighted priority updates on key features;

[0112] Based on the selected update strategy, the system calculates the normalization parameters for each feature dimension. In the sliding window update strategy, the system calculates the mean, variance, maximum, and minimum values ​​of the collected data within the window and updates the normalization parameters periodically. In the exponentially weighted moving average strategy, the system weights and merges the current and historical data according to a set smoothing coefficient, calculates the dynamic mean and variance, and updates the maximum and minimum boundaries to reflect recent distribution changes. During the adjustment process, key features are updated first according to their weights, so that their normalization parameters respond more quickly to the drift of operating conditions, thereby ensuring the consistency of the feature scale of the input prediction model. According to the feature weight matrix, key features are given priority in the normalization parameter update process. For features with higher weights, the adjustment magnitude and frequency are higher than those of ordinary features to ensure that their normalized mean, variance, and boundary values ​​more accurately reflect the latest operating conditions. The system monitors the update magnitude during the update process to prevent over-adjustment or abnormal jumps. The continuity of parameter changes is ensured by comparing with historical data. Non-key features adopt a standard update strategy to maintain the overall feature distribution stability.

[0113] The updated feature vector is output to the prediction engine module, and the history of normalization parameter adjustment is recorded at the same time.

[0114] After the normalization parameters are updated, the system applies the updated mean, variance, maximum, and minimum values ​​to the current feature vector, performs uniform scale normalization on all features, and combines the processed feature vectors into a matrix according to time order and operating condition label order. This matrix is ​​then output to the prediction engine module, including adjustment timestamps, feature dimensions, original parameter values, updated parameter values, update strategies used, and feature weight information. All historical records are stored in time series, which can be used to analyze the evolution trend of normalization parameters, drift response effects, and future optimization of normalization strategies. The recorded data can also provide a traceability basis for long-term energy consumption analysis.

[0115] Prediction Engine Module: Receives the output features from the normalized control module, uses a statistical model to predict energy consumption trends, combines the prediction results with operating condition labels for weighted correction, and outputs them to the feedback optimization module;

[0116] The process of predicting energy consumption trends using a statistical model in the prediction engine module is as follows:

[0117] Receive the feature vector output by the normalization control module;

[0118] The system obtains a multi-dimensional feature vector matrix that has undergone weighted normalization from the normalization control module. This matrix contains multi-scale statistical features and trend features divided by time, space and operating conditions. The feature vector matrix is ​​verified for integrity before transmission to ensure that no data is lost or duplicated and to maintain the feature order consistent with that during historical model training. The system also records the feature source, timestamp and corresponding operating condition label to provide data basis for subsequent prediction and deviation analysis.

[0119] Prediction models are selected based on historical energy consumption data and current characteristics, including linear regression, random forest, and deep neural networks;

[0120] The system analyzes the correlation between currently collected features and historical energy consumption data, including linear correlation coefficient, feature importance ranking, and autocorrelation analysis. Based on the analysis results, it selects the appropriate prediction model type. For energy consumption sequences with a linear trend in data distribution, a linear regression model is used; for sequences with strong nonlinear relationships, a random forest model is used; and for sequences with multidimensional feature combinations and long-term time dependence, a deep neural network model is used. The model selection process is automatically matched using statistical performance indicators such as mean squared error, mean absolute error, and coefficient of determination, and a model configuration table is generated for prediction use.

[0121] The energy consumption trend is predicted based on the trained model, and the predicted value and confidence interval are output.

[0122] After selecting a model, the system inputs the normalized feature vector into the prediction model for forward computation to generate energy consumption predictions for future time steps. For linear regression and random forest models, the system calculates the predicted value and its confidence interval for each time step and obtains the error range through variance propagation. For deep neural network models, the system uses multiple forward inferences and Monte Carlo sampling to estimate the uncertainty of the output. The prediction results include the energy consumption value, confidence interval, standard deviation, and feature contribution for each time step.

[0123] Monitor prediction deviations and abnormal fluctuations, and transmit the prediction results to the feedback optimization module;

[0124] During the model prediction process, the system simultaneously monitors the input features and output results in real time, calculates the deviation distribution between the prediction residual and historical errors, identifies the impact of abnormal fluctuations or sudden operating conditions on the prediction results, and includes anomaly detection based on standard deviation threshold judgment and time series drift monitoring. When the predicted value deviates from the historical trend by more than the set standard, the system generates anomaly markers and records the anomaly time, magnitude and corresponding feature dimensions to provide a basis for the feedback optimization module. After completing the prediction and anomaly monitoring, the system outputs the predicted energy consumption value, confidence interval, feature contribution and anomaly marker data in a structured manner to the feedback optimization module.

[0125] The weighted correction process in the prediction engine module, which combines the prediction results with the working condition labels, is as follows:

[0126] Based on the current operating condition labels and future operating condition trends, dynamic weighting coefficients are assigned to the prediction results;

[0127] The root system receives operating condition labels and future trend predictions from the operating condition perception module, maps the operating condition information corresponding to each time step to the feature vector of the prediction result, and assigns initial weight coefficients to each operating condition category through a preset weight mapping table. The weight coefficients are dynamically calculated based on the frequency of operating condition occurrence, historical energy consumption response amplitude, and future trend probability. The weight of each feature dimension is stored independently so that it can be weighted in subsequent correction steps. The weight calculation adopts a normalization method to ensure that the sum of all weights remains consistent, which is convenient for direct application to the adjustment of prediction results.

[0128] A deviation analysis was conducted between the original prediction results and historical energy consumption data, combining prediction error statistics, drift indicators, and historical deviation distribution.

[0129] The system compares the energy consumption sequence output by the prediction model with the historical energy consumption data for the corresponding time period, calculates the deviation value for each time step, including absolute error, relative error and standardized residual. At the same time, it analyzes the time series characteristics and operating condition dependence of the deviation by combining prediction error statistics, drift index and historical deviation distribution. The deviation analysis results include the deviation amplitude, trend change rate and abnormal deviation markers for each time step, providing input basis for weighted correction.

[0130] The correction results are optimized using weighted adjustment or minimum mean square error optimization methods.

[0131] Based on the deviation information obtained in step two, the system performs an initial weighted adjustment on the original prediction results according to the dynamic weights allocated in step one. At the same time, it uses the minimum mean square error optimization algorithm to iteratively correct the adjusted prediction results. The optimization process includes calculating the weighted residual sum of squares, solving for the optimal correction coefficient, and updating the prediction value at each time step. This ensures that the corrected sequence minimizes the impact of historical deviations and drift while maintaining consistency of operating conditions. The optimization iteration adopts a batch processing method to ensure that the correction results within each time window maintain temporal continuity.

[0132] The weighting coefficients, correction magnitudes, and error changes at each step of the correction process are recorded in real time, generating correction logs and deviation evolution curves. The energy consumption prediction results after dynamic weighting correction and error optimization are output to the feedback optimization module.

[0133] During the weighted correction and optimization iteration process, the system records the changes in weight coefficients, the magnitude of predicted value correction, and the changes in residuals at each time step in real time, forming a traceable data record. The recorded data is stored in a time series structure, and deviation evolution curves and weight adjustment curves are generated to facilitate subsequent analysis and parameter backtracking. The system also marks the time points where abnormal fluctuations or sudden changes in operating conditions occur, providing auxiliary information for the feedback optimization module. After completing the correction and optimization, the system outputs the final energy consumption prediction sequence and the corresponding confidence intervals, weight coefficients, residuals, and deviation evolution data in a structured manner to the feedback optimization module.

[0134] Feedback optimization module: Receives prediction results, combines drift index, operating condition prediction and historical error data to perform error analysis and deviation detection, and feeds back the analysis results to the operating condition perception module, feature adaptation module and normalization control module.

[0135] The process by which the analysis results are fed back to the operating condition sensing module, the feature adaptive module, and the normalized control module in the feedback optimization module is as follows:

[0136] Statistical analysis of prediction errors is performed, and the deviation distribution is calculated by combining drift index and historical error.

[0137] The system receives energy consumption prediction results and corresponding time series data from the prediction engine module, compares them point by point with the actual measured energy consumption values, calculates the absolute error and relative error at each time step, and evaluates the trend of error change with feature drift by combining the feature drift index provided by the normalization control module. It also compares and analyzes the data with historical error data to generate deviation statistics data including mean, variance, peak deviation and distribution density. The deviation calculation results form a multidimensional deviation matrix, including time dimension, operating condition category dimension and feature dimension, providing a precise quantitative basis for feedback from each module.

[0138] Generate deviation reports and feed them back to the working condition perception module in real time to update working condition labels and prediction trends;

[0139] The system organizes the deviation statistics obtained in step one into a structured deviation report, which includes the prediction error distribution and abnormal time period identifiers under each working condition. After receiving the report, the working condition perception module compares the deviation information with the current working condition label and the future working condition trend, identifies possible offsets or misjudgments of the working condition label, and dynamically updates the working condition category label. At the same time, the system corrects the future trend prediction results and adjusts the probability distribution of working condition changes for subsequent feature extraction and normalization calculation.

[0140] The deviation analysis results are transferred to the feature adaptive module to adjust the feature weighting coefficients and normalization parameters.

[0141] After receiving the deviation analysis results, the feature adaptive module associates the error distribution with each multi-scale feature vector and adjusts the feature importance weights and normalization parameters. Specifically, it corrects the feature weighting coefficients according to time window, spatial location, and working condition category, prioritizes adjusting key features that contribute significantly to the prediction error, and updates the mean, variance, and maximum and minimum parameters based on working condition labels and future trends. The module then forms a new input matrix from the updated feature vectors, ready to be passed to the normalization control module for further processing.

[0142] The drift index and deviation information are fed back to the normalization control module to update the normalization strategy and parameters.

[0143] After receiving the deviation analysis results, the feature adaptive module associates the error distribution with each multi-scale feature vector and adjusts the feature importance weights and normalization parameters. Specifically, it corrects the feature weighting coefficients according to time windows, spatial locations, and operating conditions, prioritizing adjustments to key features that contribute significantly to prediction errors. Simultaneously, it updates the mean, variance, and maximum / minimum parameters based on operating condition labels and future trends. The module then forms a new input matrix from the updated feature vectors. After receiving the drift index and deviation information from the feature adaptive module, the normalization control module evaluates the normalization update strategy, selecting an appropriate sliding window update or exponentially weighted moving average method to dynamically adjust the mean, variance, and maximum / minimum parameters. The module prioritizes key features and important operating condition weights, while recording the historical changes and update steps of the normalization parameters to create a traceable parameter evolution curve, ensuring that the normalization operation remains synchronized with the current operating conditions and feature states.

[0144] Data feedback and parameter transfer are performed between modules to complete the normalization parameter update and feature weighting adjustment;

[0145] The system establishes data channels between modules to achieve closed-loop data interaction between the operating condition sensing module, the feature adaptive module, and the normalized control module. Deviation analysis, feature weighting coefficients, and normalized parameters are transmitted and updated synchronously between modules, forming a continuously iterative data processing link. The updated feature vectors and normalized parameters are output to the prediction engine module for the next round of energy consumption trend prediction, ensuring the consistency and traceability of input data for each module. At the same time, the timestamps, parameter values, and processing status of each step are recorded, providing a complete data foundation for long-term operation and model iteration.

[0146] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based energy consumption data analysis and prediction system, characterized by, Comprise: Working condition perception module: collect the environmental parameters and running state of the equipment, use the working condition recognition algorithm combining incremental clustering and historical pattern matching, dynamically identify the current system working condition, output the working condition label and future working condition prediction result to the feature self-adaptation module; Feature self-adaptation module: receive the working condition label and future working condition prediction result, divide and extract multi-scale features according to time, space and working condition dimensions, and perform weighted normalization processing according to the working condition label and future working condition prediction result, and the normalization parameter can be dynamically adjusted with the change of working condition, and output the processed features to the normalization control module; Normalization control module: receive the features output by the feature self-adaptation module, monitor the feature distribution drift, calculate the drift index by comparing the drift threshold with the historical distribution, select the optimal normalization update strategy according to the drift index, dynamically update the normalization parameter, and use the updated normalization parameter to re-normalize the received features, and output the re-normalized features to the prediction engine module; Prediction engine module: receive the re-normalized features output by the normalization control module, use statistical model to predict energy consumption trend, and combine the energy consumption trend prediction result with the working condition label to perform weighted correction, and output the corrected energy consumption prediction result to the feedback optimization module; Feedback optimization module: receive the corrected energy consumption prediction result, combine the drift index, future working condition prediction result and historical error data to perform error analysis and deviation detection, and feed back the analysis result to the working condition perception module, feature self-adaptation module and normalization control module.

2. The AI-based energy consumption data analysis and prediction system of claim 1, wherein, The process of dynamically identifying the current system working condition is: Collecting real-time running parameters and environmental data of the equipment, including temperature, humidity, load, current and voltage; Incremental clustering and similarity matching of current data and historical working condition patterns; Output the current working condition label and future working condition trend prediction; Monitor the sudden changes of working condition, including abnormal load fluctuation and environmental disturbance, and dynamically correct the working condition label, and real-time transfer the recognition result and prediction information to the feature self-adaptation module.

3. The AI-based energy consumption data analysis and prediction system of claim 2, wherein, The process of dividing and extracting multi-scale features according to time, space and working condition dimensions is: Multi-dimensional slicing of collected data according to time window, equipment distribution and working condition label; Calculate statistical features in each slice, including mean, variance, peak change rate and time series trend features; Calculate statistical features and trend features in each time-space slice, and weight and aggregate according to working condition weight, forming a feature matrix that can be directly input into the prediction model; Classify and label features under different working conditions, and output feature vectors of each scale to the feature self-adaptation module for weighted normalization.

4. The AI-based energy consumption data analysis and prediction system of claim 3, wherein, The process of weighted normalization processing according to working condition label and future working condition prediction result is: Assign weight coefficients related to working condition label and future working condition prediction result to each multi-scale feature vector; Scale adjustment of features using normalization function; Dynamically adjust the normalization weight according to the future working condition prediction result; Real-time record the change trend of normalization parameter, and output the weighted and normalized features to the normalization control module.

5. The AI-based energy consumption data analysis and prediction system of claim 4, wherein, The process of dynamically adjusting the normalization parameter with the change of working condition is: Real-time monitor the change amplitude and drift trend of feature distribution; Adjusting the normalization parameters, including the values of mean, variance, maximum and minimum, according to the current working condition label and the future working condition prediction result output by the working condition perception module; Continuously updating the normalization parameters using the sliding window or exponential weighted moving average method; Logging and tracking the normalization parameter updating process to form a parameter evolution curve, and outputting the updated normalization parameters to the normalization control module.

6. The AI-based energy consumption data analysis and prediction system of claim 5, wherein, The process of calculating the drift index by comparing the drift threshold with the historical distribution is as follows: Statistical analysis of the weighted and normalized feature vectors output by the feature self-adaptation module to calculate the deviation of the current distribution from the historical distribution; Comparing with the set drift threshold to generate the drift index; Evaluating the degree of drift impact in combination with feature importance and working condition label; When the drift index exceeds the threshold, triggering the normalization parameter updating strategy selection, and transferring the drift index and analysis result to the normalization control module.

7. The AI-based energy consumption data analysis and prediction system of claim 6, wherein, The process of dynamically updating the normalization parameters is as follows: According to the feature drift amplitude and feature importance, selecting the sliding window updating strategy or the exponential weighted moving average strategy to dynamically adjust the normalization parameters; Adjusting the mean, variance and maximum and minimum parameters to prioritize the weighted update of key features; Outputting the updated feature vector to the prediction engine module, while recording the history of normalization parameter adjustment.

8. The AI-based energy consumption data analysis and prediction system of claim 7, wherein, The process of energy consumption trend prediction using statistical models is as follows: Receiving the re-normalized feature vector output by the normalization control module; Selecting the applicable prediction model from the pre-set prediction models, including linear regression model, random forest model and deep neural network model, according to the historical energy consumption data and current features; Predicting the energy consumption trend based on the trained model and outputting the prediction value and confidence interval; Monitoring the prediction deviation and abnormal fluctuations, and transferring the energy consumption trend prediction result to the feedback optimization module.

9. The AI-based energy consumption data analysis and prediction system of claim 8, wherein, The process of weighted correction of energy consumption trend prediction result combined with working condition label is as follows: According to the current working condition label and the future working condition prediction result, assigning a dynamic weight coefficient to the energy consumption trend prediction result; Analyzing the deviation of the original prediction result from the historical energy consumption data to calculate the prediction error statistics, drift index and historical deviation distribution; Adjusting the original prediction result by weighting based on the deviation analysis result to obtain the preliminary correction result; Optimizing the preliminary correction result using the least mean square error optimization method; Real-time recording of the weight coefficients, correction amplitude and error changes in the correction process to generate correction logs and deviation evolution curves, and outputting the energy consumption prediction result after dynamic weighted correction and error optimization to the feedback optimization module.

10. The AI-based energy consumption data analysis and prediction system of claim 9, wherein, The process of feeding back the analysis result to the working condition perception module, feature self-adaptation module and normalization control module is as follows: Statistical analysis of the prediction error, combined with the drift index and historical error to calculate the deviation distribution; Generating a deviation report and feeding back it to the working condition perception module in real time to update the working condition label and future working condition prediction result; Transferring the deviation analysis result to the feature self-adaptation module to adjust the feature weighting coefficients and normalization parameters; Feeding back the drift index and deviation information to the normalization control module to update the normalization strategy and parameters; Data feedback and parameter transmission are performed between the modules to complete normalization parameter updating and feature weighting adjustment.

Images