CN-122000862-A - Intelligent prediction method, system, equipment and storage medium for industrial electricity

CN122000862ACN 122000862 ACN122000862 ACN 122000862ACN-122000862-A

Abstract

The invention relates to the technical field of data processing, in particular to an intelligent prediction method, system, equipment and storage medium for industrial electricity, which comprises the steps of collecting multi-source data of electricity load, production operation, environment and time calendar, fusing, cleaning and abnormal value processing, constructing a preprocessing time sequence data set, and generating state marking characteristics in the abnormal processing; the method comprises the steps of carrying out feature engineering on a data set, extracting hysteresis, sliding statistics, trend, periodicity and external causal features, combining the standardized digital features with category features to construct a model input feature set, inputting the set into a pre-trained hybrid prediction model, fusing output results of a time sequence deep learning model and an integrated learning model, and finally outputting point prediction and interval prediction of industrial electric loads. The invention obviously improves the cognition and prediction precision of the complex mode of industrial electricity through multisource fusion and refined feature engineering.

Inventors

JIANG YANSEN
LI WEI
XIE XINGCHANG
LIU FENG
Sun Erguo

Assignees

山东浪潮智慧能源科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251205

Claims (10)

1. An intelligent prediction method for industrial electricity is characterized by comprising the following steps: Collecting multi-source heterogeneous data, and performing data fusion, cleaning and outlier processing on the multi-source heterogeneous data to construct a preprocessed time sequence data set, wherein the outlier processing comprises generating an outlier state marking characteristic; performing feature engineering on the preprocessed time sequence data set, and extracting multiple types of features including hysteresis features, sliding statistics features, trend features, periodicity features and external causal features; Carrying out standardization processing on numerical type features in the multiple types of features, and constructing a feature set for model input by combining the type features; Inputting the feature set into a pre-trained hybrid prediction model, fusing output prediction results of a time sequence deep learning model and an integrated learning model in the hybrid prediction model, and outputting industrial electric load point prediction and interval prediction based on the fusion results; The method comprises the steps of training a hybrid prediction model based on a feature set generated in a historical period by taking an electric load as a tag to obtain a pre-trained hybrid prediction model, wherein the hybrid prediction model comprises a time sequence deep learning model and an integrated learning model; The multi-source heterogeneous data includes electricity load data, production operation data, environment data, and time calendar data.
2. The method of claim 1, wherein performing data fusion, cleansing and outlier processing on the multi-source heterogeneous data to construct a preprocessed time series data set comprises: the data of non-uniform sampling is aligned to the same time granularity as the electricity load data by interpolation or aggregation method; Processing missing data, wherein for continuous missing caused by equipment overhaul, creating event marking characteristics and filling by adopting a mean value of data in front and back time periods or interpolation based on production conditions; And correcting the identified abnormal value into a smooth value based on normal data at the front and rear moments or marking the abnormal value as an abnormal state characteristic.
3. The method of claim 2, wherein identifying outlier data points in conjunction with the business rule-based decision comprises: First-level identification, performing preliminary screening by adopting rules based on statistical control, wherein the rules comprise: a static threshold rule that compares the electrical load with a fixed threshold based on rated parameters of the device; A dynamic threshold rule, namely generating a dynamic threshold based on statistics of historical contemporaneous data, wherein the dynamic threshold is determined by a 3-Sigma rule or a quantile method; and identifying a second level, namely verifying the primary screening result by adopting a rule based on business logic, wherein the rule comprises the following steps: judging whether the equipment state data is consistent with the electricity load data logically or not; And (3) a mode abnormality rule, namely comparing the current power load curve with a history contemporaneous typical curve, and judging whether the morphological difference exceeds an allowable range.
4. The method of claim 1, wherein feature engineering the preprocessed time series data set comprises: the basic time sequence feature is constructed by extracting power load values of 1 hour, 3 hours, 6 hours, 12 hours, 24 hours and 168 hours before the current moment to generate a hysteresis feature, calculating the mean value, standard deviation, maximum value and minimum value of the power load in a time window of 6 hours and 24 hours before the current moment to generate a sliding statistical feature, and carrying out linear regression on the power load sequence in a preset time window before the current moment to take the obtained gradient as a trend feature. The periodic characteristics are constructed by converting time-related fields in the time sequence data set after preprocessing into cyclic coding characteristics through sine transformation and cosine transformation, and generating Boolean type characteristics for marking holidays and the day before the holidays through judging date types. The external causal feature is constructed by directly extracting or calculating the planned output, the comprehensive efficiency of equipment and the number of start-up production lines by inquiring production operation data aligned with the electricity load data, calculating the refrigerating degree and the heating degree of the temperature in the environment data based on a preset basic temperature, and generating corresponding event marking features by identifying equipment maintenance schedules and special production task work orders.
5. The method of claim 1, wherein normalizing the numeric features in the plurality of classes of features and constructing a feature set for model input in combination with the class-type features comprises: scaling the numerical type features by adopting a standardized or normalized method; splicing the scaled numerical type features and the non-scaled category type features to form a unified feature vector; the feature vectors are associated with electrical load tag values for corresponding time stamps to construct a feature set for model training and prediction.
6. The method of claim 1, wherein inputting the feature set into a pre-trained hybrid prediction model and fusing output prediction results of a time-series deep learning model and an ensemble learning model in the hybrid prediction model comprises: Simultaneously inputting the feature set into a time sequence deep learning model and an integrated learning model in the pre-trained hybrid prediction model; acquiring a first prediction result output by the time sequence deep learning model and a second prediction result output by the integrated learning model; and fusing the first prediction result and the second prediction result, wherein the fusion mode is to input the first prediction result and the second prediction result serving as input features into a meta learner to perform final prediction.
7. The method of claim 6, wherein the hybrid predictive model is constructed and trained as follows: the model architecture is characterized in that the time sequence deep learning model is a time sequence convolutional network (TCN) model or a Transformer model, and the integrated learning model is a LightGBM model; the training strategy is to independently train the time sequence deep learning model and the integrated learning model by adopting a data set strictly divided according to time sequence so as to prevent future information leakage; A loss function, namely, a Huber loss function or a quantile loss function is adopted when the time sequence deep learning model is trained; And (3) fusing the results, namely fusing the prediction results output by the two models in a mode of using a linear regression model as a meta learner.
8. An intelligent prediction system for industrial electricity, comprising: The data acquisition module is used for acquiring multi-source heterogeneous data, and carrying out data fusion, cleaning and outlier processing on the multi-source heterogeneous data to construct a preprocessed time sequence data set, wherein the outlier processing comprises generation of an outlier state marking characteristic; the feature extraction module is used for carrying out feature engineering on the preprocessed time sequence data set and extracting multiple types of features including hysteresis features, sliding statistical features, trend features, periodic features and external causal features; the feature processing module is used for carrying out standardization processing on the numerical type features in the multiple types of features and constructing a feature set for model input by combining the type features; The model prediction module is used for inputting the characteristic set into a pre-trained hybrid prediction model, fusing output prediction results of a time sequence deep learning model and an integrated learning model in the hybrid prediction model, and outputting industrial electric load point prediction and interval prediction based on the fusion results; The method comprises the steps of training a hybrid prediction model based on a feature set generated in a historical period by taking an electric load as a tag to obtain a pre-trained hybrid prediction model, wherein the hybrid prediction model comprises a time sequence deep learning model and an integrated learning model; The multi-source heterogeneous data includes electricity load data, production operation data, environment data, and time calendar data.
9. An industrial electricity intelligent prediction device, comprising: The memory is used for storing an intelligent prediction program of industrial electricity; A processor for implementing the steps of the industrial electricity intelligent prediction method according to any one of claims 1-7 when executing the industrial electricity intelligent prediction program.
10. A computer readable storage medium storing a computer program, characterized in that the readable storage medium has stored thereon an industrial electricity intelligent prediction program, which when executed by a processor, implements the steps of the industrial electricity intelligent prediction method according to any one of claims 1-7.

Description

Intelligent prediction method, system, equipment and storage medium for industrial electricity Technical Field The invention belongs to the technical field of data processing, and particularly relates to an intelligent prediction method, system, equipment and storage medium for industrial electricity. Background Industrial electric load prediction is a core link of enterprise energy management and cost control. However, the electricity consumption data in the industrial scene has the characteristics of high volatility and nonlinearity, and is influenced by the complexity of multisource factors such as production planning, ambient temperature, equipment state and the like, so that the accurate prediction difficulty is extremely high. Traditional single prediction models, such as statistical models or shallow machine learning models, often have difficulty effectively capturing long-term dependencies and complex external causal effects. In addition, the existing method generally only depends on historical load data, key causal variables such as production, environment and the like cannot be systematically fused, and abnormal working conditions in the data are insufficiently processed, so that prediction accuracy is limited, robustness is poor, and fine electricity cost optimization and abnormal early warning decision are difficult to support. Therefore, there is an urgent need for an intelligent prediction method that can integrate multisource information, quantify causal effects, and provide high-precision section prediction capability. Disclosure of Invention The invention provides an intelligent prediction method, system, equipment and storage medium for industrial electricity, which aims at solving the technical problems. In a first aspect, the present invention provides an intelligent prediction method for industrial electricity, including: Collecting multi-source heterogeneous data, and performing data fusion, cleaning and outlier processing on the multi-source heterogeneous data to construct a preprocessed time sequence data set, wherein the outlier processing comprises generating an outlier state marking characteristic; performing feature engineering on the preprocessed time sequence data set, and extracting multiple types of features including hysteresis features, sliding statistics features, trend features, periodicity features and external causal features; Carrying out standardization processing on numerical type features in the multiple types of features, and constructing a feature set for model input by combining the type features; Inputting the feature set into a pre-trained hybrid prediction model, fusing output prediction results of a time sequence deep learning model and an integrated learning model in the hybrid prediction model, and outputting industrial electric load point prediction and interval prediction based on the fusion results; The method comprises the steps of training a hybrid prediction model based on a feature set generated in a historical period by taking an electric load as a tag to obtain a pre-trained hybrid prediction model, wherein the hybrid prediction model comprises a time sequence deep learning model and an integrated learning model; The multi-source heterogeneous data includes electricity load data, production operation data, environment data, and time calendar data. In an alternative embodiment, the data fusion, cleaning and outlier processing are performed on the multi-source heterogeneous data to construct a preprocessed time series data set, including: the data of non-uniform sampling is aligned to the same time granularity as the electricity load data by interpolation or aggregation method; Processing missing data, wherein for continuous missing caused by equipment overhaul, creating event marking characteristics and filling by adopting a mean value of data in front and back time periods or interpolation based on production conditions; And correcting the identified abnormal value into a smooth value based on normal data at the front and rear moments or marking the abnormal value as an abnormal state characteristic. In an alternative embodiment, in connection with business rule based decisions, identifying outlier data points includes: First-level identification, performing preliminary screening by adopting rules based on statistical control, wherein the rules comprise: a static threshold rule that compares the electrical load with a fixed threshold based on rated parameters of the device; A dynamic threshold rule, namely generating a dynamic threshold based on statistics of historical contemporaneous data, wherein the dynamic threshold is determined by a 3-Sigma rule or a quantile method; and identifying a second level, namely verifying the primary screening result by adopting a rule based on business logic, wherein the rule comprises the following steps: judging whether the equipment state data is consistent with the electricity load data logically or not; And (