A multi-model cooperative offshore oilfield development platform energy consumption prediction method

Abstract

This invention provides a multi-model collaborative method for predicting energy consumption of offshore oilfield development platforms, belonging to the technical field of energy consumption prediction for offshore oilfield development platforms. This invention collects and preprocesses historical energy and meteorological data from offshore oilfield development platforms, extracts multi-scale time-delay feature sets, identifies the optimal time-delay window using cross-correlation analysis, and performs time-delay offset processing on meteorological features. It trains a total energy input prediction model and an air release ratio prediction model based on a gradient boosting decision tree algorithm, introduces an isolated forest model for anomaly detection, and feeds back abnormal state information to the prediction model for result optimization. Based on the prediction results and meteorological fluctuation judgment, it makes alarm judgments and generates operational suggestions. An incremental learning update strategy is used to apply elastic weight constraints to the model parameters to achieve continuous model optimization, thus solving the technical problem of insufficient accuracy in predicting energy consumption of offshore oilfield development platforms.

Description

Technical Field

[0001] This invention belongs to the field of energy consumption prediction technology for offshore oilfield development platforms. Specifically, it relates to a multi-model collaborative method for predicting energy consumption of offshore oilfield development platforms. Background Technology

[0002] Energy consumption forecasting for offshore oilfield development platforms is a crucial technical aspect of petroleum engineering. Traditional techniques primarily employ statistical analysis methods based on historical energy data or single forecasting models to predict energy consumption, estimating future energy demand by establishing regression relationships between energy consumption and time series data. However, existing technologies suffer from limitations. Energy consumption on offshore oilfield development platforms is complexly influenced by multiple factors, including changes in weather conditions, fluctuations in equipment operating status, and adjustments to production conditions. Single forecasting models struggle to fully capture the multi-scale, time-varying characteristics of energy consumption. Furthermore, current methods lack accurate modeling of the time-lag correlation between meteorological and energy data and fail to effectively integrate anomaly detection mechanisms with forecast optimization processes. In other words, existing technologies suffer from insufficient accuracy in predicting energy consumption for offshore oilfield development platforms. Summary of the Invention

[0003] In view of this, the present invention provides a multi-model collaborative method for predicting energy consumption of offshore oilfield development platforms, which can solve the technical problem of insufficient accuracy in predicting energy consumption of offshore oilfield development platforms in the prior art.

[0004] This invention is implemented as follows: This invention provides a multi-model collaborative method for predicting energy consumption of offshore oilfield development platforms, comprising the following steps: collecting historical energy data and historical meteorological data of the offshore oilfield development platform; cleaning, unifying units, and imputing missing values ​​in the historical energy data and historical meteorological data to obtain a preprocessed dataset; extracting time features, energy features, unit energy consumption features, lag features, rolling statistical features, and meteorological features from the preprocessed dataset; performing cross-correlation analysis on historical meteorological data and historical energy data to identify the optimal time lag window; performing time lag offset processing on the meteorological features based on the optimal time lag window to construct a multi-scale time lag feature set; dividing the multi-scale time lag feature set into a training set and a test set; training a total energy input prediction model, an air release ratio prediction model, and an anomaly detection model based on the training set; and then using real-time collected energy data... After feature extraction and time-lag offset processing, real-time energy data and meteorological data are input into the total energy input prediction model and the air release ratio prediction model, which output the predicted values ​​of total energy input and air release ratio for the next 1 to 7 days. After feature extraction and time-lag offset processing, real-time energy data and real-time meteorological data are input into the anomaly detection model to calculate anomaly scores. When the anomaly score is less than the anomaly threshold, it is marked as an anomaly, and the anomaly status information is fed back to the total energy input prediction model and the air release ratio prediction model to optimize the prediction results and obtain the final prediction results. Based on the predicted values ​​of total energy input, air release ratio, and anomaly scores, combined with the meteorological fluctuation judgment results, an alarm judgment is made. When the predicted values ​​of total energy input exceed the dynamic high energy consumption threshold, the predicted values ​​of air release ratio exceed the air release threshold, or the anomaly scores are less than the anomaly threshold for two consecutive days, an alarm is triggered and operation suggestions are generated.

[0005] Historical energy data includes historical fuel gas consumption, historical venting gas consumption, historical diesel consumption, historical crude oil production, historical purchased electricity consumption, and historical freshwater consumption; historical meteorological data includes historical temperature, historical wind speed, historical wind direction, historical humidity, and historical precipitation.

[0006] This includes continuously collecting new energy data and new meteorological data during model operation to construct incremental training samples, mixing the incremental training samples with seasonal samples and rare operating condition samples in the historical key sample library to form an incremental update dataset, and applying elastic weight constraints to the parameters of the total energy input prediction model and the air emission ratio prediction model for incremental learning and updating.

[0007] Among them, data cleaning identifies abnormal data points in historical energy data and historical meteorological data that deviate from the normal range, judges and corrects abnormal data points by comparing them with historical data, and deletes duplicate data records.

[0008] For missing value imputation, linear interpolation is used to estimate missing values ​​based on data from adjacent time points for time series data, while the K-nearest neighbor algorithm is used to predict and impute missing values ​​for non-time series data.

[0009] Among them, energy characteristics include total fuel consumption and fuel utilization rate. Total fuel consumption is obtained by adding historical fuel gas consumption, historical diesel consumption and historical venting gas consumption.

[0010] In the cross-correlation analysis, historical temperature was selected as the meteorological reference sequence, total fuel data was selected as the energy reference sequence, and the cross-correlation coefficient was calculated every hour within the time lag window from 0 hours to 24 hours. The maximum value of the cross-correlation coefficient and its corresponding time lag hours were recorded as the optimal time lag window.

[0011] Among them, the time-delay offset processing shifts the average temperature forward by the corresponding number of hours according to the optimal time-delay window, and shifts the maximum wind speed forward by the corresponding number of hours according to the optimal time-delay window of the maximum wind speed and total fuel quantity data.

[0012] The total energy input prediction model is structured as a gradient boosting decision tree regression model based on the LightGBM algorithm. The gradient boosting decision tree regression model contains multiple decision trees, and each decision tree outputs a predicted value through a weighted summation of its leaf nodes.

[0013] Among them, the total energy input prediction model training calculates the learning rate adjustment coefficient and dynamically adjusts the learning rate based on the learning rate adjustment coefficient. The learning rate adjustment coefficient is determined through the comprehensive fitness index.

[0014] The comprehensive fitness index is calculated by weighting the normalized training error, normalized test error, and normalized meteorological fluctuations. The weight of the normalized training error is 0.4, the weight of the normalized test error is 0.4, and the weight of the normalized meteorological fluctuations is 0.2.

[0015] The structure of the air release ratio prediction model is a gradient boosting decision tree regression model built on the LightGBM algorithm. The air release ratio is calculated by dividing the historical air release consumption by the sum of the historical fuel gas consumption and the historical air release consumption, and then multiplying by 100%.

[0016] The anomaly detection model is structured as an isolated forest model based on the isolated forest algorithm. The isolated forest model contains multiple isolated trees, and each isolated tree isolates data points by randomly selecting features and split points.

[0017] In the prediction result optimization, when abnormal state information is fed back to the total energy input prediction model, the Euclidean distance between the abnormal state characteristics and the historical normal state characteristics is calculated. When the Euclidean distance is greater than the distance threshold, the total energy input prediction value is multiplied by the abnormal correction coefficient to obtain the optimized total energy input prediction value.

[0018] The dynamic high energy consumption threshold is dynamically adjusted based on the meteorological fluctuation judgment results. When the meteorological fluctuation judgment results are significant fluctuations, the dynamic high energy consumption threshold is set to 1.2 times the historical average total energy input.

[0019] Incremental learning updates the parameters of the total energy input prediction model and the air emission ratio prediction model by calculating elastic weights. The elastic weights are obtained by calculating the squared gradient of the prediction error of each parameter with respect to historical key samples. During training, the loss function consists of a prediction error term and an elastic weight regularization term.

[0020] This invention proposes a multi-model collaborative method for predicting energy consumption of offshore oilfield development platforms. It identifies the optimal time-lag window for the impact of meteorological data on energy data through cross-correlation analysis and performs time-lag offset processing, constructing a comprehensive feature set containing multi-scale time-lag characteristics. This ensures precise temporal alignment between meteorological and energy features, accurately depicting the time-lag impact of meteorological changes on energy consumption. The invention employs a collaborative total energy input prediction model and an air release ratio prediction model, combined with an anomaly detection model to monitor data anomalies in real time and feed the anomaly information back to the prediction model for result optimization, realizing a dynamic correction mechanism in the prediction process. Furthermore, this invention uses an incremental learning update strategy and applies elastic weight constraints to the model parameters, preserving the predictive ability of historical key samples while introducing new data, thus preventing the model from forgetting historical patterns. In summary, this invention solves the technical problem of insufficient accuracy in predicting energy consumption of offshore oilfield development platforms mentioned in the background art. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a graph showing the change in the cross-correlation coefficient between meteorological and energy data over a time lag window.

[0023] Figure 3 A comparison chart of the predicted total energy input for the next 7 days and the dynamic high energy consumption threshold. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0025] like Figure 1The diagram shown is a flowchart of a multi-model collaborative method for predicting energy consumption of offshore oilfield development platforms provided by this invention. This method includes the following steps:

[0026] S1. Collect historical energy data and historical meteorological data from offshore oilfield development platforms. Clean the historical energy data and historical meteorological data, unify units, and fill in missing values ​​to obtain a preprocessed dataset.

[0027] S2. Extract time features, energy features, unit energy consumption features, lag features, rolling statistical features and meteorological features from the preprocessed dataset. Perform cross-correlation analysis on historical meteorological data and historical energy data to identify the optimal time lag window. Perform time lag offset processing on meteorological features based on the optimal time lag window to construct a multi-scale time lag feature set.

[0028] S3. Divide the multi-scale time delay feature set into a training set and a test set, and train the total energy input prediction model, the air release ratio prediction model and the anomaly detection model based on the training set respectively.

[0029] S4. After the real-time energy data and real-time meteorological data collected in real time are processed by the same feature extraction and time delay offset as in step S2, they are input into the total energy input prediction model and the air release ratio prediction model, and the total energy input prediction value and air release ratio prediction value for the next 1 to 7 days are output.

[0030] S5. After feature extraction and time lag offset processing, real-time energy data and real-time meteorological data are input into the anomaly detection model to calculate the anomaly score. When the anomaly score is less than the anomaly threshold, it is marked as an anomaly state and the anomaly state information is fed back to the total energy input prediction model and the air release ratio prediction model to optimize the prediction results.

[0031] S6. Based on the total energy input forecast, the air release ratio forecast, and the anomaly score, combined with the meteorological fluctuation judgment results, an alarm is triggered and operation suggestions are generated when the total energy input forecast exceeds the dynamic high energy consumption threshold, the air release ratio forecast exceeds the air release threshold, or the anomaly score is less than the anomaly threshold for two consecutive days.

[0032] S7. During the model operation, continuously collect new energy data and new meteorological data to construct incremental training samples. Mix the incremental training samples with seasonal samples and rare operating condition samples in the historical key sample library to form an incremental update dataset. Apply elastic weight constraints to the parameters of the total energy input prediction model and the air release ratio prediction model for incremental learning and updating.

[0033] The historical energy data includes historical fuel gas consumption, historical vented air consumption, historical diesel consumption, historical crude oil production, historical purchased electricity consumption, and historical freshwater consumption. The historical meteorological data includes historical temperature, historical wind speed, historical wind direction, historical humidity, and historical precipitation. The real-time energy data includes real-time fuel gas consumption, real-time vented air consumption, real-time diesel consumption, real-time crude oil production, real-time purchased electricity consumption, and real-time freshwater consumption. The real-time meteorological data includes real-time temperature, real-time wind speed, real-time wind direction, real-time humidity, and real-time precipitation.

[0034] The data cleaning steps specifically include: identifying anomalous data points in historical energy and meteorological data that deviate from the normal range; determining and correcting anomalous data points by comparing them with historical data; and deleting duplicate data records. The unit standardization step specifically includes: converting historical fuel gas consumption into a standardized unit. Historical diesel consumption was uniformly converted to L, and historical purchased electricity consumption was uniformly converted to L. Historical freshwater consumption and historical crude oil production are uniformly converted to tons (t). The missing value imputation steps specifically include: estimating missing values ​​for time-series data using linear interpolation based on data from adjacent time points; and predicting and imputing missing values ​​for non-time-series data using the K-nearest neighbor algorithm.

[0035] The time characteristics include yearly, monthly, daily, quarterly, and winter / summer identification characteristics. The energy characteristics include total fuel consumption and fuel utilization rate. Total fuel consumption is obtained by adding historical fuel gas consumption, historical diesel consumption, and historical venting gas consumption. Fuel utilization rate is obtained as the ratio of effectively utilized fuel to total fuel consumption. The unit energy consumption characteristics include fuel consumption per unit of crude oil production and freshwater consumption per unit of crude oil production. Fuel consumption per unit of crude oil production is obtained by dividing total fuel consumption by historical crude oil production, and freshwater consumption per unit of crude oil production is obtained by dividing historical freshwater consumption by historical crude oil production.

[0036] The lag features include 1-day, 2-day, 3-day, 7-day, and 14-day lag features, each corresponding to the historical energy and meteorological data values ​​before the corresponding number of days. The rolling statistical features include 3-day rolling mean, 3-day rolling standard deviation, 7-day rolling mean, 7-day rolling standard deviation, 14-day rolling mean, and 14-day rolling standard deviation. The meteorological features include average temperature, maximum wind speed, average humidity, and cumulative precipitation.

[0037] The specific steps of the cross-correlation analysis include: selecting historical temperature as a meteorological reference sequence, selecting total fuel consumption data as an energy reference sequence, calculating the cross-correlation coefficient every hour within a time lag window from 0 to 24 hours, and recording the maximum value of the cross-correlation coefficient and its corresponding time lag hours as the optimal time lag window. The specific steps of the time lag offset processing include: shifting the average temperature forward by the corresponding number of hours according to the optimal time lag window, and shifting the maximum wind speed forward by the corresponding number of hours according to the optimal time lag window of the maximum wind speed and total fuel consumption data, so that the shifted meteorological characteristics are aligned with the current energy characteristics in time.

[0038] The total energy input prediction model is structured as a gradient boosting decision tree regression model based on the LightGBM algorithm. This model comprises multiple decision trees, each outputting a predicted value through a weighted summation of its leaf nodes. The final total energy input prediction value is obtained by accumulating the predicted values ​​from all decision trees. The steps for establishing the training dataset for the total energy input prediction model specifically include: dividing the multi-scale time-delay feature set according to time order, selecting the first 70% of the data as the training set, and the last 30% as the test set. Each sample in the training set contains all features of the multi-scale time-delay feature set as input features, and the total energy input at the corresponding time point as the label value. The total energy input is obtained by adding historical fuel gas consumption, historical vented gas consumption, historical diesel consumption, historical crude oil production, historical purchased electricity consumption, and historical freshwater consumption.

[0039] The specific steps for training the total energy input prediction model include: setting an initial learning rate of 0.05, a number of decision trees of 200, a maximum depth of 8, and a minimum number of samples per leaf node of 20; iteratively training the decision trees on the training set using the gradient boosting method, calculating the residual between the current model prediction value and the true label value in each iteration, constructing a new decision tree to fit the residual, and adding the new decision tree to the gradient boosting decision tree regression model; calculating the learning rate adjustment coefficient and dynamically adjusting the learning rate according to the learning rate adjustment coefficient; calculating the root mean square error and mean absolute error on the test set to evaluate the model performance, and completing the training when the root mean square error is less than 5% of the mean total energy input.

[0040] The calculation steps for the learning rate adjustment coefficient specifically include: extracting the training set prediction error, test set prediction error, and meteorological characteristic fluctuation value for the current training round; the training set prediction error is calculated by averaging the absolute values ​​of the differences between the current model prediction value and the true label value on the training set; the test set prediction error is calculated by averaging the absolute values ​​of the differences between the current model prediction value and the true label value on the test set; the meteorological characteristic fluctuation value is calculated by summing the standard deviation of the average temperature and the standard deviation of the maximum wind speed in the training set for the current training round; calculating the normalized training error, normalized test error, and normalized meteorological fluctuation; the normalized training error is calculated by dividing the training set prediction error by the average total energy input; the normalized test error is calculated by dividing the test set prediction error by the average total energy input; the normalized meteorological fluctuation is calculated by dividing the meteorological characteristic fluctuation value by the historical meteorological characteristic fluctuation standard value, wherein the historical... The standard value for meteorological characteristic fluctuation is the average of the sum of the standard deviations of average temperature and maximum wind speed in historical meteorological data. The comprehensive fitness index is calculated by weighting the normalized training error, normalized test error, and normalized meteorological fluctuation. The weight of normalized training error is 0.4, the weight of normalized test error is 0.4, and the weight of normalized meteorological fluctuation is 0.2. The learning rate adjustment coefficient is determined based on the range of the comprehensive fitness index: 1.2 when the comprehensive fitness index α∈[0, 0.3), 1.0 when the comprehensive fitness index α∈[0.3, 0.6), 0.8 when the comprehensive fitness index α∈[0.6, 1.0), and 0.5 when the comprehensive fitness index α≥1.0. The initial learning rate is multiplied by the learning rate adjustment coefficient to obtain the dynamically adjusted learning rate.

[0041] The structure of the air release ratio prediction model is a gradient boosting decision tree regression model built based on the LightGBM algorithm. This model contains multiple decision trees, each outputting a predicted value through a weighted sum of its leaf nodes. The final predicted air release ratio is obtained by accumulating the predicted values ​​from all decision trees. The steps for establishing the training dataset for the air release ratio prediction model specifically include: dividing the multi-scale time-delay feature set according to time order, selecting the first 70% of the data as the training set, and the last 30% as the test set. Each sample in the training set contains all features of the multi-scale time-delay feature set as input features, and the corresponding air release ratio as the label value. The air release ratio is calculated by dividing the historical air release consumption by the sum of the historical fuel gas consumption and the historical air release consumption, and then multiplying by 100%.

[0042] The specific steps for training the air release ratio prediction model include: setting an initial learning rate of 0.05, a decision tree count of 200, a maximum depth of 8, and a minimum number of samples per leaf node of 20; iteratively training the decision trees on the training set using the gradient boosting method, calculating the residual between the current model prediction value and the true label value in each iteration, constructing a new decision tree to fit the residual, and adding the new decision tree to the gradient boosting decision tree regression model; calculating the learning rate adjustment coefficient and dynamically adjusting the learning rate according to the learning rate adjustment coefficient; calculating the root mean square error and mean absolute error on the test set to evaluate the model performance, and completing the training when the root mean square error is less than 10% of the average air release ratio.

[0043] The anomaly detection model is structured as an isolated forest model based on the isolated forest algorithm. This model comprises multiple isolated trees, each of which isolates data points by randomly selecting features and split points. Anomaly scores are calculated based on the path length of each data point within the isolated tree, and the average of all isolated tree anomaly scores is used as the final anomaly score. The steps for establishing the training dataset for the anomaly detection model specifically include: using all data from the multi-scale time-delay feature set as training data, and using all features from the multi-scale time-delay feature set as input features for each sample.

[0044] The specific steps for training the anomaly detection model include: setting the number of isolated trees to 100, with each isolated tree having a sample ratio of 0.8; randomly selecting 80% of the samples from the training data for each isolated tree, randomly selecting a feature and a split point of the feature in the samples, dividing the samples into two groups, and recursively constructing isolated trees until each leaf node contains only one sample or reaches the maximum depth; calculating the average path length of each sample in the training data across all isolated trees, normalizing the average path length into an anomaly score, with the anomaly score ranging from -1 to 0, and the smaller the value, the higher the degree of anomaly.

[0045] The steps for optimizing the prediction results specifically include: when abnormal state information is fed back to the total energy input prediction model, extracting the energy data features and meteorological data features at the time corresponding to the abnormal state, calculating the Euclidean distance between the abnormal state features and the historical normal state features, determining a true anomaly when the Euclidean distance is greater than a distance threshold, and multiplying the total energy input prediction value by an anomaly correction coefficient to obtain the optimized total energy input prediction value; when abnormal state information is fed back to the air release ratio prediction model, extracting the real-time air release consumption and real-time fuel gas consumption at the time corresponding to the abnormal state, calculating the deviation rate between the air release ratio under the abnormal state and the average air release ratio under the historical normal state, determining a true anomaly when the deviation rate is greater than a deviation threshold, and adding the anomaly correction value to the air release ratio prediction value to obtain the optimized air release ratio prediction value.

[0046] The anomaly correction coefficient is calculated by dividing the absolute value of the difference between the anomaly score and the anomaly threshold by the absolute value of the anomaly threshold and then adding 1. The anomaly correction value is calculated by multiplying the deviation rate by the average proportion of air release. The distance threshold is set to twice the standard deviation of the Euclidean distance of the historical normal state characteristics. The deviation threshold is set to 20%.

[0047] The specific steps for determining the meteorological fluctuation result include: calculating the difference between the real-time temperature and the historical temperature at the same time the previous day as the temperature fluctuation value, and calculating the difference between the real-time wind speed and the historical wind speed at the same time the previous day as the wind speed fluctuation value; when the absolute value of the temperature fluctuation value is greater than the temperature fluctuation threshold or the absolute value of the wind speed fluctuation value is greater than the wind speed fluctuation threshold, it is determined that the meteorological data has significant fluctuations, and the meteorological fluctuation determination result is significant fluctuation; when the absolute value of the temperature fluctuation value is less than or equal to the temperature fluctuation threshold and the absolute value of the wind speed fluctuation value is less than or equal to the wind speed fluctuation threshold, it is determined that the meteorological data does not have significant fluctuations, and the meteorological fluctuation determination result is no significant fluctuation.

[0048] The dynamic high-energy consumption threshold is dynamically adjusted based on the meteorological fluctuation judgment results. When the meteorological fluctuation judgment result is significant, the dynamic high-energy consumption threshold is set to 1.2 times the historical average total energy input; when the meteorological fluctuation judgment result is no significant fluctuation, the dynamic high-energy consumption threshold is set to 1.1 times the historical average total energy input. The air release threshold is set to 8%. The abnormal threshold is set to -0.5. The temperature fluctuation threshold is set to 5℃. The wind speed fluctuation threshold is set to 3m / s.

[0049] The steps for generating the operation suggestions specifically include: when a high energy consumption alarm is triggered, querying the high energy consumption response strategies in the knowledge base; if the real-time temperature is below 10℃, generating an operation suggestion to increase insulation measures; if the real-time temperature is above 30℃, generating an operation suggestion to adjust the power of the cooling equipment; if the real-time temperature is between 10℃ and 30℃, generating an operation suggestion to reduce the gas valve opening by 5%; when an air release alarm is triggered, querying the air release response strategies in the knowledge base; if the real-time wind speed is less than 2m / s, generating an operation suggestion to start the recovery compressor; if the real-time wind speed is greater than or equal to 2m / s, generating an operation suggestion to increase air release diffusion monitoring; when an abnormal alarm is triggered, querying the abnormal response strategies in the knowledge base, generating an operation suggestion to check whether the instruments and equipment are working properly.

[0050] The construction steps of the historical key sample library specifically include: extracting seasonal samples and rare operating condition samples from historical training data. The seasonal samples include typical energy consumption data for each season, such as spring, summer, autumn, and winter. The 10% of data with the largest energy consumption fluctuations in each season are selected as seasonal samples. The rare operating condition samples include abnormal data with anomaly scores less than -0.3 and extreme data with total energy input exceeding 1.5 times the historical average total energy input. The seasonal samples and rare operating condition samples are merged to form the historical key sample library.

[0051] The steps for constructing the incremental training samples specifically include: collecting new energy data and new meteorological data from the past 7 days; performing the same data cleaning, unit unification, and missing value imputation processing on the new energy data and new meteorological data as in step S1; performing the same feature extraction and time lag offset processing on the processed new energy data and new meteorological data as in step S2; and using the processed data as incremental training samples.

[0052] The incremental learning and updating steps specifically include: mixing incremental training samples with the historical key sample library at a ratio of 3:1 to form an incremental update dataset; calculating elastic weights for the parameters of the total energy input prediction model and the air emission ratio prediction model, wherein the elastic weights are obtained by calculating the squared gradient of the prediction error of each parameter with respect to the historical key samples; training the total energy input prediction model and the air emission ratio prediction model on the incremental update dataset, wherein the loss function during training consists of a prediction error term and an elastic weight regularization term, wherein the prediction error term is calculated by the mean square error between the predicted value and the true value, and the elastic weight regularization term is calculated by summing the product of the parameter update amount and the elastic weight, and the total loss function is the weighted sum of the prediction error term and the elastic weight regularization term; setting the weight of the prediction error term to 0.7 and the weight of the elastic weight regularization term to 0.3, and iteratively training the model parameters for 50 rounds on the incremental update dataset.

[0053] The comprehensive fitness index is used to evaluate the overall performance of the total energy input prediction model and the air emission ratio prediction model during training, comprehensively considering the impact of training error, testing error, and meteorological fluctuations on model performance. The learning rate adjustment coefficient is a coefficient used to dynamically adjust the learning rate during the training of the total energy input prediction model and the air emission ratio prediction model, with different adjustment factors determined according to different value ranges of the comprehensive fitness index. The normalized training error is the value after normalizing the prediction error of the training set, eliminating the influence of dimensions for easier calculation with other indicators. The normalized testing error is the value after normalizing the prediction error of the test set, eliminating the influence of dimensions for easier calculation with other indicators. The normalized meteorological fluctuation is the value after normalizing the meteorological characteristic fluctuation value, eliminating the influence of dimensions for easier calculation with other indicators.

[0054] The optimal time lag window is the best time delay for the impact of meteorological data on energy data, identified through cross-correlation analysis, reflecting the time lag effect of meteorological changes on energy consumption. The multi-scale time lag feature set is a comprehensive feature set including time features, energy features, unit energy consumption features, lag features, rolling statistical features, and meteorological features processed by time lag offset, capturing the changing patterns of energy data across multiple time scales and the time lag impact of meteorological data. The abnormal state feature is a combination of energy data features and meteorological data features at the time corresponding to the abnormal state. The historical normal state feature is a combination of energy data features and meteorological data features corresponding to normal states with an anomaly score greater than or equal to -0.3 in historical data.

[0055] The elastic weights are weighting coefficients used to constrain the parameter updates of the total energy input prediction model and the air release ratio prediction model. They reflect the importance of the parameters to the prediction performance of historical key samples; parameters with higher importance are subject to stronger constraints during incremental learning. The parameter update amount is the change in model parameters relative to their values ​​before the update during incremental learning. The knowledge base is a database storing response strategies based on actual production experience and professional knowledge, including high energy consumption response strategies, air release response strategies, and anomaly response strategies.

[0056] Optionally, the present invention also provides a computer-based method for forming a multi-model collaborative energy consumption prediction system for offshore oilfield development platforms. The computer is equipped with a readable storage medium that stores program instructions, which execute the above-described method when the computer is run.

[0057] The specific implementation methods of the above steps are described in detail below.

[0058] The specific implementation of step S1 involves real-time collection of historical energy and meteorological data using sensors and metering equipment deployed at key locations on the offshore oilfield development platform. Historical energy data includes historical fuel gas consumption, historical venting gas consumption, historical diesel consumption, historical crude oil production, historical purchased electricity consumption, and historical freshwater consumption. Historical meteorological data includes historical temperature, historical wind speed, historical wind direction, historical humidity, and historical precipitation. The data cleaning process uses statistical analysis to identify outlier data points deviating from the normal range. A three-standard-deviation criterion is used to determine if data points are abnormal; data points exceeding the mean plus or minus three standard deviations are marked as outliers. Corrections are made by comparing with historical data, and duplicate data records are deleted to ensure data uniqueness. The unit unification process converts historical fuel gas consumption to a standardized value. Historical diesel consumption is uniformly converted to L, and historical purchased electricity consumption is uniformly converted to Historical freshwater consumption and historical crude oil production were uniformly converted to tons (t) to provide a standardized data foundation for subsequent analysis. For the missing value imputation process, linear interpolation was used for time series data, estimating the missing value based on data values ​​at adjacent time points. For non-time series data, the K-nearest neighbor algorithm was used, finding the K nearest neighbor data points most similar in characteristics to the missing data point and taking their mean as the predicted imputation value. The K-value was set to 5 as a reference.

[0059] Step S2 specifically involves multi-dimensional feature extraction and time-lag relationship analysis of the preprocessed dataset. Time feature extraction directly obtains yearly, monthly, and daily features from the data timestamps. Quarterly features are defined based on the month, and winter / summer identifiers are determined based on seasonal information: winter is from December to February of the following year, and summer is from June to August. In energy feature calculation, total fuel consumption is obtained by the arithmetic sum of historical fuel gas consumption, historical diesel consumption, and historical venting gas consumption. Fuel utilization rate is obtained by the ratio of effective utilization to total fuel consumption, where effective utilization is the total fuel consumption minus venting gas consumption. In unit energy consumption feature generation, fuel consumption per unit of crude oil production and freshwater consumption per unit of crude oil production are calculated by dividing total fuel consumption and historical freshwater consumption by historical crude oil production, respectively, eliminating the interference of production fluctuations on energy consumption analysis. Lag feature introduction involves extracting historical energy data and historical meteorological data from 1, 2, 3, 7, and 14 days ago as lag features for the current moment, capturing the historical state of energy data at different time scales. Rolling statistical feature calculations are performed by calculating the rolling mean and rolling standard deviation of historical energy data for the past 3, 7, and 14 days, respectively. The rolling mean reflects the average level of the data within the time window, while the rolling standard deviation reflects the degree of data fluctuation. Meteorological feature extraction includes average temperature, maximum wind speed, average humidity, and cumulative precipitation. Cross-correlation analysis uses the Pearson correlation coefficient to calculate the correlation between historical temperature and total fuel consumption data under different time lag windows. The correlation coefficient is calculated every hour from 0 to 24 hours, and the time lag hour corresponding to the maximum correlation coefficient is identified as the optimal time lag window. This optimal time lag window reflects the time delay effect of meteorological changes on energy consumption. Time lag offset processing shifts the average temperature and maximum wind speed forward by the corresponding number of hours based on the optimal time lag window, aligning meteorological features with energy features in time. This constructs a multi-scale time lag feature set that includes time features, energy features, unit energy consumption features, lag features, rolling statistical features, and shifted meteorological features.

[0060] The specific implementation of step S3 involves dividing the multi-scale time-delay feature set into a training set and a test set in a 70% to 30% ratio according to time sequence, and training the total energy input prediction model, the air release ratio prediction model, and the anomaly detection model respectively. Both the total energy input prediction model and the air release ratio prediction model employ a gradient boosting decision tree regression model based on the LightGBM algorithm. This model achieves nonlinear regression prediction by integrating multiple decision trees. Each decision tree fits the residual of the previous model during training, and the loss function is minimized using the gradient descent method. The initial learning rate is set to 0.05, the number of decision trees is 200, the maximum depth is 8, and the minimum number of samples per leaf node is 20. These parameter settings balance model complexity and generalization ability. The learning rate adjustment coefficient is calculated by extracting the prediction error of the training set, the prediction error of the test set, and the meteorological characteristic fluctuation value. These three are divided by their respective standardized benchmark values ​​to obtain the normalized training error, normalized test error, and normalized meteorological fluctuation. These are then weighted and summed with weights of 0.4, 0.4, and 0.2 to obtain the comprehensive fitness index. The learning rate adjustment coefficient is determined based on the range of the comprehensive fitness index. When the comprehensive fitness index is less than 0.3, the learning rate adjustment coefficient is 1.2, indicating good model convergence and faster training. When the comprehensive fitness index is between 0.3 and 0.6, the learning rate adjustment coefficient is 1.0, maintaining the original learning rate. When the comprehensive fitness index is between 0.6 and 1.0, the learning rate adjustment coefficient is 0.8, reducing the learning rate to prevent overfitting. When the comprehensive fitness index is greater than or equal to 1.0, the learning rate adjustment coefficient is 0.5, significantly reducing the learning rate and stabilizing training. The anomaly detection model adopts the isolated forest model based on the isolated forest algorithm. The isolated forest model constructs multiple isolated trees by randomly selecting features and split points. It calculates the anomaly score by utilizing the characteristic that abnormal data points are easier to isolate. The anomaly score is obtained by normalizing the average path length of the data point in the isolated tree. The shorter the path length, the easier it is to isolate, i.e., the higher the degree of anomaly.

[0061] The specific implementation of step S4 is to process the real-time energy data and real-time meteorological data collected in real time through the same data cleaning, unit unification and missing value completion processes as in step S1, and then through the same feature extraction and time delay offset processes as in step S2, to obtain a real-time multi-scale time delay feature set. The real-time multi-scale time delay feature set is then input into the trained total energy input prediction model and air emission ratio prediction model. The total energy input prediction value and air emission ratio prediction value for the next 1 to 7 days are output through the weighted summation of multiple decision trees in the model, providing a prediction basis for energy supply planning and air emission management.

[0062] The specific implementation of step S5 involves inputting the real-time multi-scale time-delay feature set into the anomaly detection model. An anomaly score is obtained by calculating the average path length of real-time data points in the isolated tree. When the anomaly score is less than the anomaly threshold of -0.5, it is marked as an anomaly state. The anomaly state information is fed back to the total energy input prediction model and the air release ratio prediction model for prediction result optimization. In the prediction result optimization process, energy data features and meteorological data features corresponding to the anomaly state are extracted to form anomaly state features. The Euclidean distance between the anomaly state features and historical normal state features is calculated. When the Euclidean distance is greater than the distance threshold, it is determined to be a genuine anomaly rather than data fluctuation. The total energy input prediction value is multiplied by an anomaly correction coefficient to obtain the optimized total energy input prediction value. The anomaly correction coefficient is calculated by dividing the absolute value of the difference between the anomaly score and the anomaly threshold by the absolute value of the anomaly threshold and then adding 1. The optimization of the air release ratio prediction model is achieved by calculating the deviation rate between the air release ratio under abnormal conditions and the average air release ratio under historical normal conditions. When the deviation rate is greater than the deviation threshold of 20%, it is determined to be a real anomaly. The optimized air release ratio prediction value is obtained by adding the anomaly correction value to the predicted air release ratio value. The anomaly correction value is calculated by multiplying the deviation rate by the average air release ratio.

[0063] The specific implementation of step S6 involves alarm judgment based on the predicted total energy input, the predicted air release ratio, and the anomaly score, combined with the meteorological fluctuation judgment results. Meteorological fluctuation judgment is achieved by calculating the difference between the real-time temperature and the historical temperature at the same time the previous day, resulting in a temperature fluctuation value; and by calculating the difference between the real-time wind speed and the historical wind speed at the same time the previous day, resulting in a wind speed fluctuation value. Significant fluctuations are judged when the absolute value of the temperature fluctuation value exceeds the temperature fluctuation threshold by 5°C or the absolute value of the wind speed fluctuation value exceeds the wind speed fluctuation threshold by 3 m / s. The dynamic high energy consumption threshold is adjusted according to the meteorological fluctuation judgment results. It is set to 1.2 times the historical average total energy input for significant fluctuations and 1.1 times for no significant fluctuations. This dynamic adjustment mechanism avoids false alarms caused by changes in meteorological conditions. Alarm triggering adopts a two-day consecutive rule. An alarm is triggered when the predicted total energy input exceeds the dynamic high energy consumption threshold, the predicted air release ratio exceeds the air release threshold by 8% for two consecutive days, or the anomaly score is less than the anomaly threshold -0.5. This consecutive rule effectively filters false alarms caused by instantaneous fluctuations. Operation suggestion generation involves querying the knowledge base to match the current alarm type and weather conditions, and generating targeted operation suggestions based on real-time temperature and wind speed.

[0064] The specific implementation of step S7 involves continuously collecting new energy data and new meteorological data during model operation. After the same preprocessing and feature extraction as steps S1 and S2, incremental training samples are constructed. Seasonal samples and rare operating condition samples are extracted from historical training data to construct a historical key sample library. The incremental training samples and the historical key sample library are mixed at a 3:1 ratio to form an incremental update dataset. Elastic weight calculation involves calculating the gradient of the prediction error of each parameter of the total energy input prediction model and the air emission ratio prediction model on the historical key samples. The square of the gradient is used as the elastic weight, which reflects the importance of the parameter to the prediction performance of the historical key samples. The incremental learning and update process trains the model on the incremental update dataset. The loss function consists of a prediction error term and an elastic weight regularization term. The prediction error term is calculated using the mean squared error, and the elastic weight regularization term is calculated by summing the product of the parameter update amount and the elastic weight. The total loss function is weighted and summed with weights of 0.7 and 0.3. The elastic weight regularization term constrains the update magnitude of important parameters to prevent new data from overwriting historical key knowledge. The model update is completed after 50 iterations of training.

[0065] It should be noted that the key technical ideas of this invention include multi-scale time-delay feature construction, adaptive learning rate adjustment, and elastic weight constraint incremental learning. Multi-scale time-delay feature construction identifies the optimal time-delay window for meteorological and energy data through cross-correlation analysis, and performs time-delay offset processing on meteorological features to solve the problem of complex time delays in the impact of meteorological changes on energy consumption. Compared with traditional synchronous feature methods, the multi-scale time-delay features can accurately capture the delayed impact of temperature drops on heating energy demand 6 to 12 hours later, significantly improving the contribution of meteorological features to the prediction model, enabling the model to predict energy demand trends in advance before meteorological conditions change. Adaptive learning rate adjustment evaluates the combined impact of training error, testing error, and meteorological fluctuations through a comprehensive fitness index. The learning rate is dynamically adjusted according to the interval of the comprehensive fitness index. Compared with fixed learning rate training methods, this adaptive adjustment mechanism accelerates training when the model converges well and reduces the learning rate to stabilize training when the model shows signs of overfitting or severe meteorological fluctuations. This allows the model to adapt to the complex and variable meteorological conditions and production environment of offshore oil fields, improving model training efficiency and prediction stability. Elastic weight-constrained incremental learning uses the squared gradient of the prediction error of parameters on historical key samples as elastic weights. During incremental learning, regularization constraints are applied to important parameters. Compared to methods that directly update the model using new data, this elastic weight constraint mechanism protects parameters crucial for identifying seasonal patterns and rare operating conditions from being overwritten by new data, solving the catastrophic forgetting problem in online incremental learning. This allows the model to retain its historical key knowledge while continuously learning new data. The synergistic effect of these three technical approaches lies in the fact that multi-scale time-lag features provide the model with inputs that accurately reflect the coupling relationship between meteorology and energy. Adaptive learning rate adjustment ensures that the model can efficiently and stably learn this coupling relationship during training. Elastic weight-constrained incremental learning allows the model to continuously optimize during operation without losing learned key patterns. Together, these three elements construct an intelligent prediction system that can accurately predict energy demand, adapt to environmental changes, and possess continuous learning capabilities. Compared to traditional static prediction methods, this system exhibits higher prediction accuracy, stronger environmental adaptability, and better long-term stability.

[0066] It should be noted that this invention also solves the following technical problem: Existing technologies lack a continuous optimization mechanism for energy consumption prediction models of offshore oilfield development platforms, leading to a gradual degradation in predictive performance when faced with new operating conditions during long-term operation. This invention constructs a historical key sample library to store seasonal and rare operating condition samples, mixes incremental training samples with historical key samples in a proportional manner to form an incremental update dataset, calculates elastic weights for model parameters, and applies elastic weight regularization constraints during incremental learning. This allows the model to maintain its predictive ability for historical key operating conditions while learning new data patterns, avoiding the catastrophic forgetting phenomenon caused by new data overwriting historical knowledge in traditional incremental learning methods, and achieving long-term stability of model predictive performance. Furthermore, this invention also solves the technical problem of insufficient reliability of prediction results under abnormal energy consumption conditions. It uses an isolated forest model to calculate anomaly scores in real time and determine abnormal states, extracts features of abnormal states and compares them with historical normal state features using Euclidean distance to verify the authenticity of the anomalies, and applies a correction coefficient based on the degree of anomaly to the prediction results of real abnormal operating conditions, ensuring that the prediction model can still output reasonable prediction values ​​under abnormal conditions such as equipment failure and sudden changes in operating conditions.

[0067] Specifically, the principle of this invention is as follows: This invention solves the technical problem of insufficient accuracy in predicting energy consumption of offshore oilfield development platforms. Its principle lies in quantifying the time-delay correlation strength between meteorological parameters and energy consumption through cross-correlation analysis, determining the time delay in the impact of meteorological changes on energy consumption, and then offsetting meteorological features according to the optimal time-delay window, eliminating feature correlation distortion caused by time misalignment. This allows the prediction model to learn energy consumption patterns based on accurately aligned temporal relationships. This invention employs a gradient boosting decision tree regression model to construct a total energy input prediction model and an air release ratio prediction model. By calculating a comprehensive fitness index, the learning rate is dynamically adjusted, enabling the model training process to adaptively optimize the convergence speed based on the combined performance of training error, testing error, and meteorological fluctuations. This invention introduces an isolated forest model for anomaly detection and feeds back anomaly information to the prediction model. By calculating the Euclidean distance between anomaly features and historical normal features, true anomalies are identified and correction coefficients are applied, achieving targeted compensation for prediction deviations under abnormal operating conditions. This aligns with the physical logic that energy system consumption patterns deviate from normal distribution under abnormal conditions.

[0068] The following provides a specific embodiment 1 of the present invention. The specific implementation of the calculation steps involved in this embodiment 1 is described in detail below.

[0069] In step S2, the specific implementation methods for time characteristics, energy characteristics, and unit energy consumption characteristics are as follows. Total fuel quantity The calculation formula is expressed as follows:

[0070] ;

[0071] In the formula, Total fuel quantity, in units of ; Historical fuel gas consumption, in units of ; Historical diesel consumption is converted into equivalent fuel gas volume, in units of ; Historical air consumption, in units of .

[0072] in, The conversion formula is expressed as follows:

[0073] ;

[0074] In the formula, This represents diesel fuel consumption, expressed in liters (L). This refers to the energy density of diesel fuel, which is 36 MJ / L by default. This refers to the calorific value of the fuel gas, which is 38 MJ / L by default. .

[0075] Fuel utilization rate The calculation formula is expressed as follows:

[0076] ;

[0077] In the formula, Fuel utilization rate, dimensionless; The effective utilization amount is expressed in units of... .

[0078] Fuel consumption per unit of crude oil production The calculation formula is expressed as follows:

[0079] ;

[0080] In the formula, Fuel consumption per unit of crude oil production, dimensionless; Historical crude oil production, in tons (t). The baseline fuel quantity is 10,000 (empirical value). ; The benchmark crude oil production is 100 tons, based on empirical values.

[0081] Freshwater consumption per unit of crude oil production The calculation formula is expressed as follows:

[0082] ;

[0083] In the formula, Freshwater consumption per unit of crude oil production, dimensionless; Historical freshwater consumption, in tons (t). The baseline freshwater consumption is 50 tons, based on empirical values.

[0084] The specific implementation method of cross-correlation analysis is as follows. Cross-correlation coefficient. The calculation formula is expressed as follows:

[0085] ;

[0086] In the formula, Time delay The cross-correlation coefficient corresponding to the hour is dimensionless. The time lag is in hours, ranging from 0 to 24 hours; For the first Historical temperature for the hour, in °C; This represents the historical average temperature, expressed in °C. For the first Total fuel consumption per hour, in units ; This is the average of the total fuel consumption, in units of... ; The total number of hours in the historical data. Optimal time lag window. correspond The time delay in hours when the maximum value is taken.

[0087] In step S3, the specific implementation method for establishing the training datasets for the total energy input prediction model and the air release ratio prediction model is as follows. Total Energy Input The calculation formula is expressed as follows:

[0088] ;

[0089] In the formula, Total energy input, in units of ; Historical purchased electricity consumption converted into equivalent fuel gas volume, unit: ; Historical freshwater consumption is converted into equivalent fuel gas volume, in units of .

[0090] in, The conversion formula is expressed as follows:

[0091] ;

[0092] In the formula, Historical external electricity consumption, in kW·h; This is the energy conversion factor, with a default value of 0.3. / kW·h.

[0093] The conversion formula is expressed as follows:

[0094] ;

[0095] In the formula, Historical freshwater consumption, in tons (t). This is the freshwater energy consumption coefficient, which is 2 by default. / t.

[0096] Air release ratio The calculation formula is expressed as follows:

[0097] ;

[0098] In the formula, The percentage of air released is dimensionless.

[0099] The specific implementation method for calculating the learning rate adjustment coefficient is as follows. Training set prediction error. The calculation formula is expressed as follows:

[0100] ;

[0101] In the formula, The training set prediction error is expressed in the same unit as the label value. The number of samples in the training set; For the first Predicted values ​​for each training sample; For the first The true label value of each training sample.

[0102] Test set prediction error The calculation formula is expressed as follows:

[0103] ;

[0104] In the formula, The test set prediction error is expressed in the same unit as the label value. This represents the number of samples in the test set. For the first Predicted values ​​for each test sample; For the first The true label value of each test sample.

[0105] Meteorological characteristic fluctuation value The calculation formula is expressed as follows:

[0106] ;

[0107] In the formula, These are meteorological characteristic fluctuation values, expressed in °C + m / s. This represents the standard deviation of the average temperature during the current training session, in °C. This represents the standard deviation of the maximum wind speed in the training set for the current training round, in m / s.

[0108] Normalized training error The calculation formula is expressed as follows:

[0109] ;

[0110] In the formula, To normalize the training error, dimensionless; The average total energy input, in units of same.

[0111] Normalized test error The calculation formula is expressed as follows:

[0112] ;

[0113] In the formula, This is a normalized test error, dimensionless.

[0114] Normalized meteorological fluctuations The calculation formula is expressed as follows:

[0115] ;

[0116] In the formula, Normalized meteorological fluctuations, dimensionless; This represents the standard value for historical meteorological characteristic fluctuations, with units equal to... same.

[0117] Historical meteorological characteristic fluctuation standard value The calculation formula is expressed as follows:

[0118] ;

[0119] In the formula, This represents the standard value for historical meteorological characteristic fluctuations, with units of ℃+m / s; This represents the number of historical statistical periods, typically taken as the number of months in the historical data. For the first The standard deviation of the average temperature over a period of time, in °C; For the first The standard deviation of the maximum wind speed within each period is expressed in m / s.

[0120] Comprehensive fitness index The calculation formula is expressed as follows:

[0121] ;

[0122] In the formula, It is a comprehensive fitness index, dimensionless.

[0123] Learning rate adjustment factor Based on the comprehensive fitness index The value of is determined when hour ,when hour ,when hour ,when hour Dynamically adjusted learning rate The calculation formula is expressed as follows:

[0124] ;

[0125] In the formula, The learning rate is dynamically adjusted and is dimensionless. This is the initial learning rate, typically set to 0.05.

[0126] In step S5, the specific implementation method for optimizing the prediction results is as follows: Euclidean distance. The calculation formula is expressed as follows:

[0127] ;

[0128] In the formula, The distance is Euclidean and dimensionless. The number of feature dimensions; For abnormal state number One eigenvalue; For the historical normal state The average of the features; For the first The standard deviation of each feature is used for normalization.

[0129] Distance threshold The calculation formula is expressed as follows:

[0130] ;

[0131] In the formula, The distance threshold is dimensionless. The standard deviation of the Euclidean distance between historical normal state characteristics is obtained by calculating the standard deviation of the pairwise Euclidean distance between historical normal state samples.

[0132] Anomaly Correction Coefficient The calculation formula is expressed as follows:

[0133] ;

[0134] In the formula, This is an anomaly correction coefficient, dimensionless; These are outlier scores, ranging from -1 to 0. This is the abnormal threshold, which defaults to -0.5.

[0135] Deviation rate The calculation formula is expressed as follows:

[0136] ;

[0137] In the formula, The deviation rate is dimensionless. The percentage of air released under abnormal conditions is dimensionless. The percentage of air released under normal historical conditions is the average, and is dimensionless.

[0138] Abnormal correction value The calculation formula is expressed as follows:

[0139] ;

[0140] In the formula, This is an anomaly correction value, dimensionless.

[0141] In step S6, the specific implementation method for determining the meteorological fluctuation result is as follows: Temperature fluctuation value. The calculation formula is expressed as follows:

[0142] ;

[0143] In the formula, This represents temperature fluctuation values, expressed in °C. This is the real-time temperature, in °C. The temperature is the historical temperature at the same time the previous day, in °C.

[0144] Wind speed fluctuation value The calculation formula is expressed as follows:

[0145] ;

[0146] In the formula, This represents wind speed fluctuation values, in m / s. This is real-time wind speed, in m / s. The wind speed at the same time the previous day is the historical wind speed in m / s.

[0147] when or The time was determined to be significant fluctuations in meteorological data, among which This is the temperature fluctuation threshold, with a default value of 5℃. This is the wind speed fluctuation threshold, with a default value of 3 m / s. Dynamic high energy consumption threshold. Adjustments are made based on the results of meteorological fluctuation assessments when significant fluctuations are observed in meteorological data. When meteorological data does not fluctuate significantly The specific implementation method for alarm detection is the same as described above, and will not be repeated in detail here.

[0148] In step S7, the specific implementation method for incremental learning and updating is as follows: Elastic weights. The calculation formula is expressed as follows:

[0149] ;

[0150] In the formula, For parameters The elastic weights are dimensionless. This refers to the number of key historical samples. For the first The prediction error of a key sample, dimensionless; For the first One model parameter; For prediction error For parameters The gradient.

[0151] Prediction error term The calculation formula is expressed as follows:

[0152] ;

[0153] In the formula, This is the prediction error term, with the same unit as the square of the label value; To incrementally update the number of samples in the dataset; For the first Predicted values ​​for each incremental sample; For the first The true value of each incremental sample.

[0154] Total loss function The calculation formula is expressed as follows:

[0155] ;

[0156] In the formula, The total loss function is dimensionless. To normalize the prediction error, the variance of the training set label values ​​is used, with units equal to... same; This represents the total number of model parameters. For the first time before the update Each parameter value; The normalized baseline value for the elastic weights is dimensionless, and the calculation formula is as follows: .

[0157] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0158] To verify the effectiveness of this invention, technicians set up a test environment and collected historical operating data from an offshore oilfield development platform for 365 days, from January to December 2024. Historical energy data included historical fuel gas consumption, historical venting gas consumption, historical diesel consumption, historical crude oil production, historical purchased electricity consumption, and historical freshwater consumption. Historical meteorological data included historical temperature, historical wind speed, historical wind direction, historical humidity, and historical precipitation. Technicians performed data cleaning on the collected historical energy and meteorological data, identifying 87 abnormal data points. These were corrected by comparing with historical data from adjacent periods, and 12 duplicate records were deleted. The historical fuel gas consumption was uniformly converted to [a more accurate and concise version]. Historical diesel consumption is uniformly converted to L, and historical purchased electricity consumption is uniformly converted to Historical freshwater consumption and historical crude oil production were uniformly converted to tons (t). For missing value imputation, linear interpolation was used to estimate the missing points in the 34 time-series data points, and the K-nearest neighbor algorithm was used to predict and impute the missing points in the 19 non-time-series data points, resulting in a complete preprocessed dataset.

[0159] Technicians extracted time features, energy features, unit energy consumption features, lag features, rolling statistical features, and meteorological features from the preprocessed dataset. Time features included yearly, monthly, daily, quarterly, and winter / summer indicator features. The total fuel consumption in the energy features was obtained by adding the daily historical fuel gas consumption, historical diesel consumption, and historical venting gas consumption; the total fuel consumption for a typical day was 156,320. Fuel utilization rate, calculated as the ratio of effective utilization to total fuel consumption, was 92.3% for that day. Fuel consumption per unit of crude oil production, a characteristic of unit energy consumption, was obtained by dividing total fuel consumption by historical crude oil production; the fuel consumption per unit of crude oil production for that day was 8.74. The freshwater consumption per unit of crude oil production is calculated by dividing historical freshwater consumption by historical crude oil production. The daily freshwater consumption per unit of crude oil production is 0.52 t / t. Lag characteristics include 1-day, 2-day, 3-day, 7-day, and 14-day lag characteristics. Rolling statistical characteristics include 3-day rolling mean, 3-day rolling standard deviation, 7-day rolling mean, 7-day rolling standard deviation, 14-day rolling mean, and 14-day rolling standard deviation. Meteorological characteristics include average temperature, maximum wind speed, average humidity, and cumulative precipitation.

[0160] Technicians conducted cross-correlation analysis on historical meteorological and energy data to identify the optimal time lag window. Historical temperature was selected as the meteorological reference series, and total fuel consumption data was selected as the energy reference series. The cross-correlation coefficient was calculated for each hour within the time lag window, from 0 to 24 hours. Figure 2 As shown, the cross-correlation coefficient reaches its maximum value of 0.847 at a time lag of 6 hours, thus determining the optimal time lag window to be 6 hours. Based on this optimal time lag window, technicians performed time lag shifting on the meteorological features, shifting the average temperature forward by 6 hours and the maximum wind speed forward by 5 hours according to the optimal time lag window for the maximum wind speed and total fuel consumption data. This ensures that the shifted meteorological features are aligned with the current energy features in time, constructing a multi-scale time lag feature set containing 38 feature dimensions.

[0161] Technicians divided the multi-scale time-delay feature set according to time sequence, selecting the first 70% (256 days) of data as the training set and the last 30% (109 days) of data as the test set. Based on the training set, a total energy input prediction model was trained. The total energy input was obtained by adding historical fuel gas consumption, historical vented gas consumption, historical diesel consumption, historical crude oil production, historical purchased electricity consumption, and historical freshwater consumption. The average total energy input in the training set was 172450. The technicians set an initial learning rate of 0.05, a decision tree size of 200, a maximum depth of 8, and a minimum number of samples per leaf node of 20. They iteratively trained the decision trees on the training set using gradient boosting. At the 50th training iteration, the prediction error on the training set was 6892. The prediction error on the test set is 7235. The meteorological characteristic fluctuation value was 8.6, the normalized training error was 0.040, the normalized test error was 0.042, the normalized meteorological fluctuation was 0.073, the comprehensive fitness index was 0.047, the learning rate adjustment coefficient was determined to be 1.2, and the dynamically adjusted learning rate was 0.06. At the 120th training round, the prediction error of the training set was 5124. The prediction error on the test set was 5687. The overall fitness index was 0.035, and the learning rate adjustment coefficient was set to 1.2. The root mean square error calculated on the test set was 6234. The mean absolute error is 4876. The root mean square error is less than 5% of the mean total energy input, thus completing the training of the total energy input prediction model.

[0162] Technicians trained a venting percentage prediction model based on the training set. The venting percentage was calculated by dividing historical venting consumption by the sum of historical fuel gas consumption and historical venting consumption, then multiplying by 100%. The average venting percentage in the training set was 6.8%. Technicians trained the model with the same parameters, achieving a root mean square error (RMSE) of 0.52% and a mean absolute error (MAO) of 0.38% on the test set. The RMSE was less than 10% of the average venting percentage, thus completing the venting percentage prediction model training. Technicians also trained an anomaly detection model based on the training set, setting the number of isolated trees to 100 and the sample ratio of each isolated tree to 0.8. For each isolated tree, 80% of the samples from the training data were randomly selected for training. In the training data, 23 anomaly states with anomaly scores less than -0.5 were identified.

[0163] Technicians collected real-time energy and meteorological data for seven days, from January 1st to January 7th, 2025, for predictive testing. After feature extraction and time-delay offset processing, the real-time energy and meteorological data were input into the total energy input prediction model and the air release ratio prediction model. The prediction results for the next seven days are shown in Table 1.

[0164] Table 1. Energy Consumption Forecast for the Next 7 Days

[0165]

[0166] Technicians input real-time energy and meteorological data into an anomaly detection model after feature extraction and time lag offset processing to calculate daily anomaly scores. On January 4th, the anomaly score was -0.58, less than the anomaly threshold of -0.5, marking it as an anomaly. This anomaly information was then fed back to the total energy input prediction model for prediction result optimization. Technicians extracted energy and meteorological data features corresponding to the anomaly on January 4th, calculated the Euclidean distance between the anomaly features and historical normal features (156.8), and the distance threshold was 72.4. A distance greater than the threshold was considered a genuine anomaly. The anomaly correction coefficient was calculated by dividing the absolute value of the difference between the anomaly score and the anomaly threshold by the absolute value of the anomaly threshold and adding 1. The anomaly correction coefficient was 1.16. The predicted total energy input value was 194720. Multiplying by the anomaly correction factor yields the optimized total energy input forecast of 225875. On January 5th, the anomaly score was -0.62, which is less than the anomaly threshold. Therefore, the prediction results were optimized, and the optimized total energy input prediction value was 229348. .

[0167] Technicians analyzed real-time meteorological data to determine weather fluctuations. The real-time temperature on January 4th was -2.3℃, while the historical temperature at the same time the previous day was 4.8℃, resulting in a temperature fluctuation of -7.1℃. Since the absolute value of this fluctuation (7.1℃) exceeded the temperature fluctuation threshold of 5℃, the data was deemed to exhibit significant fluctuations. Based on this meteorological fluctuation assessment, the dynamic high-energy consumption threshold was set to the historical average total energy input of 172450. 1.2 times that of the previous year, with a dynamic high energy consumption threshold of 206940. .like Figure 3 As shown, the total energy input forecast for the two consecutive days of January 4th and January 5th is 225875. and 229348 All exceeded the dynamic high energy consumption threshold of 206940. This triggered a high energy consumption alarm. Technicians consulted the knowledge base for high energy consumption response strategies. The real-time temperature on January 4th was -2.3℃, below 10℃, and generated operational suggestions to increase insulation measures, recommending increasing the thickness of the insulation layer on the platform pipelines and activating the heat tracing system.

[0168] During model operation, technicians continuously collected new energy and meteorological data for seven days, from January 8th to January 14th, to construct incremental training samples. The new energy and meteorological data underwent data cleaning, unit standardization, and missing value imputation. Feature extraction and time-lag offset processing were then applied to the processed data, resulting in seven incremental training samples. From historical training data, technicians extracted seasonal and rare operating condition samples to construct a historical key sample library. For each season (spring, summer, autumn, and winter), the 10% of data with the largest energy consumption fluctuations were selected, yielding 37 seasonal samples. Rare operating condition samples included 23 anomalous data points with anomaly scores less than -0.3 and 18 extreme data points with total energy input exceeding 1.5 times the historical average, yielding 41 rare operating condition samples. The historical key sample library contained a total of 78 samples. The technicians mixed the seven incremental training samples with the 78 samples from the historical key sample library at a 3:1 ratio (the actual mixing ratio was 7:26), forming an incremental update dataset of 33 samples. Technicians calculated elastic weights for the parameters of the total energy input prediction model and the air release ratio prediction model. The elastic weights were obtained by calculating the squared gradient of the prediction error of each parameter with respect to historical key samples. The model was trained on the incremental update dataset. During the training process, the loss function consisted of a prediction error term and an elastic weight regularization term. The prediction error term had a weight of 0.7, and the elastic weight regularization term had a weight of 0.3. The model parameters were updated after 50 rounds of iterative training.

[0169] The advancements of this invention compared to traditional methods lie in its precise identification of the time-lag window of meteorological data's impact on energy data through cross-correlation analysis and the subsequent time-lag offset processing. This eliminates the correlation distortion problem caused by the temporal misalignment of meteorological and energy characteristics in traditional methods, enabling the prediction model to learn energy consumption patterns based on accurately aligned temporal relationships, thus significantly improving prediction accuracy. This invention employs a multi-model collaborative mechanism, using an anomaly detection model to monitor abnormal data states in real time and feed anomaly information back to the prediction model for result optimization. This solves the problem of large prediction deviations in traditional single prediction models under abnormal operating conditions, achieving targeted compensation for abnormal situations such as equipment failure and sudden changes in operating conditions. Furthermore, this invention utilizes an incremental learning update strategy and applies elastic weight constraints to model parameters. While introducing new data, it retains the predictive ability for historical key samples, overcoming the catastrophic forgetting phenomenon caused by new data overwriting historical knowledge in traditional incremental learning methods, ensuring the long-term stability of the model's predictive performance. This invention dynamically adjusts the learning rate by calculating a comprehensive fitness index, enabling the model training process to adaptively optimize the convergence speed based on the combined performance of training error, testing error, and weather fluctuations. This avoids the problem of mismatched convergence speeds at different training stages in traditional fixed learning rate methods, thereby improving model training efficiency and final performance.

[0170] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0171] Table 2. Variable Explanation Table (Part 1)

[0172]

[0173] Table 3. Variable Explanation Table (Part Two)

[0174]

[0175] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-model collaborative method for predicting energy consumption of offshore oilfield development platforms, characterized in that, Includes the following steps: Historical energy and meteorological data from offshore oilfield development platforms were collected, and the historical energy and meteorological data were cleaned, standardized in units, and missing values ​​were filled in to obtain a preprocessed dataset. Time features, energy features, unit energy consumption features, lag features, rolling statistical features, and meteorological features are extracted from the preprocessed dataset. Cross-correlation analysis is performed on historical meteorological data and historical energy data to identify the optimal time lag window. Based on the optimal time lag window, the meteorological features are subjected to time lag offset processing to construct a multi-scale time lag feature set. The multi-scale time lag feature set is divided into a training set and a test set. Based on the training set, a total energy input prediction model, an air release ratio prediction model, and an anomaly detection model are trained respectively. Real-time energy data and real-time meteorological data are collected in real time, processed by feature extraction and time lag offset processing, and then input into the total energy input prediction model and the air release ratio prediction model. The predicted values ​​of total energy input and air release ratio for the next 1 to 7 days are output. Real-time energy data and real-time meteorological data are processed by feature extraction and time lag offset, and then input into the anomaly detection model to calculate anomaly scores. When the anomaly score is less than the anomaly threshold, it is marked as an anomaly state. The anomaly state information is fed back to the total energy input prediction model and the air release ratio prediction model to optimize the prediction results and obtain the final prediction results.

2. The method according to claim 1, characterized in that, It also includes alarm judgment based on the total energy input forecast, the air release ratio forecast, and the anomaly score combined with the meteorological fluctuation judgment results. When the total energy input forecast exceeds the dynamic high energy consumption threshold, or the air release ratio forecast exceeds the air release threshold, or the anomaly score is less than the anomaly threshold for two consecutive days, an alarm is triggered and operation suggestions are generated.

3. The method according to claim 2, characterized in that, It also includes continuously collecting new energy data and new meteorological data during model operation to build incremental training samples, mixing the incremental training samples with seasonal samples and rare operating condition samples in the historical key sample library to form an incremental update dataset, and applying elastic weight constraints to the parameters of the total energy input prediction model and the air emission ratio prediction model for incremental learning and updating.

4. The method according to claim 3, characterized in that, Data cleaning identifies anomalous data points in historical energy and meteorological data that deviate from the normal range. By comparing these anomalous data points with historical data, the system corrects them and deletes duplicate data records.

5. The method according to claim 4, characterized in that, Missing value imputation uses linear interpolation to estimate missing values ​​based on data from adjacent time points for time series data, and the K-nearest neighbor algorithm is used to predict and impute missing values ​​for non-time series data.

6. The method according to claim 5, characterized in that, Energy characteristics include total fuel consumption and fuel utilization rate. Total fuel consumption is obtained by adding historical fuel gas consumption, historical diesel consumption, and historical venting gas consumption.

7. The method according to claim 6, characterized in that, Cross-correlation analysis uses historical temperature as the meteorological reference series and total fuel consumption data as the energy reference series. It calculates the cross-correlation coefficient every hour within the time lag window from 0 to 24 hours and records the maximum value of the cross-correlation coefficient and its corresponding time lag hours as the optimal time lag window.

8. The method according to claim 7, characterized in that, The time-delay offset processing shifts the average temperature forward by the corresponding number of hours according to the optimal time-delay window, and shifts the maximum wind speed forward by the corresponding number of hours according to the optimal time-delay window of the maximum wind speed and total fuel quantity data.

9. The method according to claim 8, characterized in that, The total energy input prediction model is structured as a gradient boosting decision tree regression model based on the LightGBM algorithm. The gradient boosting decision tree regression model contains multiple decision trees, and each decision tree outputs a predicted value by weighted summation of its leaf nodes.

10. The method according to claim 9, characterized in that, Historical energy data includes historical fuel gas consumption, historical vented gas consumption, historical diesel consumption, historical crude oil production, historical purchased electricity consumption, and historical freshwater consumption; historical meteorological data includes historical temperature, historical wind speed, historical wind direction, historical humidity, and historical precipitation.

Images