Low-carbon agricultural machinery park methanol power generation-power consumption collaborative scheduling system

Abstract

The application provides a low-carbon agricultural machinery park methanol power-electricity collaborative scheduling system, and belongs to the technical field of power system scheduling and control. The system comprises a multi-source data acquisition module, a data preprocessing module, a load prediction module, a methanol power unit operation modeling module and a collaborative scheduling optimization module. The application aims at the low-carbon agricultural machinery park methanol power-electricity collaborative scheduling demand, constructs a multi-module collaborative and full-link closed-loop intelligent scheduling system, and takes high-precision load prediction as the driving force, so that the optimal collaborative operation of the methanol power unit and the power grid is realized.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching and control technology, and in particular to a methanol power generation-consumption coordinated dispatching system for low-carbon agricultural machinery parks. Background Technology

[0002] As a core scenario for agricultural production, agricultural machinery parks exhibit significant cyclical fluctuations in electricity load, characterized by multi-device collaboration and spatial-temporal dispersion. During peak agricultural seasons such as sowing and harvesting, high-power equipment like harvesters and dryers operate intensively, with peak loads reaching 3-5 times the normal level. Conversely, during off-seasons, only basic electricity needs for office and storage operations are required, resulting in a sharp drop in load. The current power supply and dispatching model for these parks suffers from multiple pain points:

[0003] The energy structure is highly carbon-intensive: traditionally relying on diesel generators for emergency power supply results in high carbon emission intensity, which contradicts the dual carbon goals; some industrial parks rely solely on the power grid, resulting in high peak-hour electricity prices and power supply reliability that is greatly affected by grid fluctuations.

[0004] The application of methanol power generation is extensive: In industrial parks where methanol generator sets have been introduced, most adopt a constant power operation mode without dynamically adjusting the output according to load changes, resulting in wasted methanol consumption and frequent start-ups and shutdowns of the units, which damage the equipment.

[0005] Insufficient load forecasting accuracy: Existing forecasts mostly use a single time series algorithm (such as ARIMA), which does not integrate multi-dimensional features such as agricultural machinery operation plans and environmental factors. This makes it impossible to accurately capture short-term load fluctuations and long-term load trends, resulting in a disconnect between scheduling schemes and actual needs.

[0006] Poor dispatch coordination: The dispatching of the power generation side (methanol unit) and the power consumption side (agricultural machinery and park load) is isolated, and no coordinated mechanism of load forecasting-unit output-grid interaction has been formed, which poses a risk of power supply gap or energy waste and makes it difficult to guarantee the reliability of power supply. Summary of the Invention

[0007] This invention provides a methanol power generation and power consumption coordinated dispatch system for low-carbon agricultural machinery parks. It constructs a multi-module coordinated, full-link closed-loop intelligent dispatch system to meet the needs of methanol power generation and power consumption coordinated dispatch in low-carbon agricultural machinery parks. The core is driven by high-precision load forecasting to achieve optimal coordinated operation between methanol generator sets and the power grid.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The low-carbon agricultural machinery park methanol power generation-electricity consumption coordinated dispatch system includes:

[0010] The multi-source data acquisition module is used to collect load correlation data, agricultural machinery operation plan data, environmental data, power grid data and methanol generator set operation data of agricultural machinery parks, and output structured raw datasets;

[0011] The data preprocessing module is used to preprocess the structured raw dataset, extract short-term, medium-term, and long-term feature subsets, and output a standardized dataset.

[0012] The load forecasting module receives a standardized dataset and constructs a forecasting model using the ARIMA (Autoregressive Integral Moving Average) algorithm, the XGBoost (Extreme Gradient Boosting) algorithm, and the LSTM (Long Short-Term Memory) algorithm. The short-term forecast values ​​and residuals output by the ARIMA algorithm are used as input features for the XGBoost algorithm, and the medium-term forecast values ​​and residuals output by the XGBoost algorithm are used as input features for the LSTM algorithm. The residuals output by the LSTM algorithm are fed back to the ARIMA and XGBoost algorithms to correct their parameters. After interactive correction by the three algorithms, the load forecasting results are output through weighted fusion.

[0013] The methanol generator set operation modeling module is used to receive the unit-type data of the standardized dataset, construct the output-energy consumption model, cost model, carbon emission model and operating constraints, and output the methanol generator set operation model set;

[0014] The collaborative scheduling optimization module is used to receive the power grid data from the time-sharing load forecast result set, the methanol generator set operation model set, and the standardized dataset. With the goals of low carbon, economy, and reliability, it solves the optimal solution through a multi-objective particle swarm optimization algorithm combined with the entropy weight TOPSIS (approximation ideal solution sorting) method, and outputs the optimal collaborative scheduling scheme set.

[0015] This manual also includes the following: The methanol power generation and electricity consumption coordinated dispatch system for low-carbon agricultural machinery parks

[0016] The scheduling instruction execution and real-time monitoring module is used to receive the optimal collaborative scheduling scheme set and the time-sharing load prediction result set, convert the scheme into equipment execution instructions, collect actual operating data in real time and calculate the deviation from the planned value, adjust the instructions through the deviation correction algorithm, and output the real-time operating status dataset and abnormal alarm information.

[0017] The scheduling effect evaluation module is used to receive the real-time operation status dataset, abnormal alarm information, optimal collaborative scheduling scheme set and historical benchmark data of the standardized dataset, evaluate the effect from the dimensions of economy, low carbon, reliability and scheduling fit, and output an evaluation report containing optimization suggestions. The evaluation report is fed back to the multi-source data acquisition module, load forecasting module and collaborative scheduling optimization module respectively.

[0018] In this specification, the weighted fusion process of the load forecasting module is as follows: First, the average absolute percentage error of the ARIMA algorithm, XGBoost algorithm, and LSTM algorithm is calculated, and the initial weights of the three algorithms are allocated according to the inverse error ratio rule. Then, the interaction coefficients between the algorithms are introduced to correct the initial weights. The interaction coefficients include the interaction strength of ARIMA to XGBoost, the interaction strength of XGBoost to ARIMA, the interaction strength of XGBoost to LSTM, and the interaction strength of LSTM to the first two algorithms. The final weights after correction satisfy the condition that the sum is 1. The fused prediction value is calculated by multiplying the ARIMA-corrected prediction value by the ARIMA weight, the XGBoost-corrected prediction value by the XGBoost weight, and the LSTM prediction value by the LSTM weight.

[0019] In this specification, the preprocessing and feature extraction of the data preprocessing module specifically include: using the 3σ criterion + isolated forest algorithm to process outliers; using the 3σ criterion for data with a granularity of 15 minutes or less, and the isolated forest algorithm for data with a granularity of 1 hour or more; replacing outliers with the weighted average of the three adjacent periods, with the weight of recent data being 0.6; filling missing values ​​with linear interpolation (for high-frequency data) or similar daily replacement + trend correction (for low-frequency data); during feature extraction, the ARIMA algorithm extracts the load values ​​of the previous three periods, the first-order difference of the load, and the number of short-term start-stops of agricultural machinery at the 15-minute granularity; the XGBoost algorithm extracts the ARIMA predicted values, ARIMA residuals, the number of agricultural machinery running at the hourly level, and the historical load of the same period at the 1-hour granularity; and the LSTM algorithm extracts the XGBoost predicted values, XGBoost residuals, ARIMA predicted values, ARIMA residuals, and the load of the same period 7 days ago at the 1-day granularity. The extracted features are filtered by Pearson correlation coefficient (retaining features with an absolute correlation coefficient > 0.3) to form a three-level feature subset.

[0020] In this specification, the output-energy consumption model of the methanol generator set operation modeling module is a piecewise linear fitting model: with 50% of the rated output of the methanol generator set as the dividing point, different fitting coefficients are set for the low load segment (output ≤ 50% of rated output) and the high load segment (output > 50% of rated output); the fitting data comes from the historical data of the generator set in the standardized dataset, including historical output values ​​and corresponding methanol consumption, and the model fitting error is controlled within ±2%.

[0021] In this specification, the low-carbon, economic, and reliable objectives of the collaborative scheduling optimization module are as follows: the low-carbon objective is to minimize the sum of carbon emissions from the methanol generator set and the carbon emissions from electricity purchased from the grid; the economic objective is to minimize the sum of methanol consumption costs, maintenance costs, start-up and shutdown costs, and electricity purchase and sale costs from the grid; and the reliability objective is to ensure a load fulfillment rate of ≥99.9%. In the multi-objective particle swarm optimization algorithm, the particle dimension includes the output of the methanol generator set, the electricity purchased and sold from the grid, and the start-up and shutdown status of the unit. The inertia weight is set to 0.7 to balance the global search and local convergence capabilities. The constraints include unit output of 200kW≤P≤1200kW, ramp rate ≤100kW / minute, start-up and shutdown interval ≥30 minutes, and electricity purchased from the grid ≤500kW and electricity sold ≤300kW.

[0022] In this specification, the deviation correction algorithm between the scheduling instruction execution and the real-time monitoring module is a PI control algorithm. In the correction instruction calculation formula, the proportional coefficient Kp = 0.8 to quickly respond to deviations, and the integral coefficient Ki = 0.2 to eliminate steady-state errors. The conditions for triggering correction are: the absolute value of the deviation between the actual output of the equipment and the planned value > 5% × planned value, or the absolute value of the deviation between the actual load of the park and the time-sharing load prediction result set > 3% × predicted value.

[0023] In this specification, the optimization suggestions of the scheduling effect evaluation module specifically include: outputting prediction deviation analysis to the load forecasting module to adjust the autoregressive coefficients of the ARIMA algorithm, the number of decision trees in the XGBoost algorithm, and the number of hidden layer neurons in the LSTM algorithm; outputting the target achievement status to the collaborative scheduling optimization module to dynamically adjust the weights of low-carbon, economic, and reliable targets; and outputting deviation data source analysis to the multi-source data acquisition module to supplement the acquisition of agricultural machinery fault signals, extreme environmental parameters, and power grid voltage fluctuation data.

[0024] In this specification, the granularity of the multi-source data acquisition module is adapted to the three-level prediction requirements of the load prediction module: agricultural machinery operation data is collected every 15 seconds, office and lighting electricity consumption is collected every minute, environmental data is collected every 15 minutes, and power grid data and agricultural machinery operation plan data are updated daily; all collected data are summarized into structured raw datasets with three-level granularity of 15 minutes, 1 hour, and 1 day to ensure granularity matching with short-term, medium-term, and long-term feature subsets.

[0025] In this specification, the carbon emission model of the methanol generator set operation modeling module adopts factor fusion calculation, and the total carbon emission = carbon emission of methanol generator set + carbon emission of electricity purchased from the grid; wherein the carbon emission of methanol generator set = methanol consumption × methanol life cycle carbon emission factor, and the carbon emission of electricity purchased from the grid = electricity purchased from the grid × grid peak-valley-flat differential carbon emission factor. Both types of carbon emission factors are derived from carbon emission data in the standardized dataset.

[0026] In summary, the present invention has at least the following beneficial effects:

[0027] Low-carbon enhancement: Replace diesel power generation with methanol power generation, and reduce carbon emissions caused by ineffective power generation by combining load forecasting and precise power output scheduling logic; optimize power purchase strategy by integrating low-carbon factors of the power grid to achieve carbon emission reduction throughout the power supply chain of the park, which meets the needs of green agricultural development.

[0028] Economic optimization: The output of methanol generator units is dynamically adjusted based on the load forecast results of multiple algorithms to avoid the waste of methanol due to over-powered generators; the power purchase and sale strategy is optimized in combination with the peak-valley-flat electricity pricing of the power grid to reduce the cost of purchasing electricity during peak hours, while reducing maintenance losses caused by frequent start-ups and shutdowns of generator units, thereby reducing the overall energy operating cost of the park.

[0029] Enhanced power supply reliability: Based on high-precision load forecasting of 15 minutes to 30 days, a dual guarantee system of methanol unit main supply + grid supplementary power is constructed. Through real-time monitoring and deviation correction mechanism, it can cope with sudden start-up and shutdown of agricultural machinery, load fluctuations and other situations, and ensure that the power supply reliability rate remains stable at a high level.

[0030] Enhanced intelligence: A fully closed-loop intelligent system is formed, encompassing data acquisition, preprocessing, prediction, scheduling, execution, and evaluation. The load prediction module continuously optimizes prediction accuracy as the system operates through multi-algorithm bidirectional interaction and dynamic correction. Scheduling decisions require no manual intervention, adapting to the trend of unmanned and intelligent management in agricultural machinery parks. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the low-carbon agricultural machinery park methanol power generation-electricity consumption coordinated dispatch system involved in this invention.

[0033] Figure 2 This is a schematic diagram of the entire data flow process of the system involved in this invention.

[0034] Figure 3 This is a schematic diagram of the interactive fusion process of the three load forecasting algorithms involved in this invention. Detailed Implementation

[0035] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0036] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0037] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0038] like Figure 1 and Figure 2 As shown, this embodiment provides a low-carbon agricultural machinery park methanol power generation-electricity consumption collaborative scheduling system. The system includes seven core modules: a multi-source data acquisition module that accurately acquires multi-dimensional data on park load, agricultural machinery operation, environment, power grid, and methanol generator units, and archives the data in multiple granularities from 15 seconds to 1 day; a data preprocessing module that cleans, normalizes, and extracts hierarchical features from the raw data to support subsequent algorithm applications; a load forecasting module that uses ARIMA, XGBoost, and LSTM algorithms in pairs for bidirectional interaction, integrating short-term stable features, medium-term multi-factor correlations, and long-term nonlinear trends to output high-precision load forecasting results for all time periods; a methanol generator unit operation modeling module that constructs a quantitative model of output, energy consumption, cost, and carbon emissions, along with physical constraints; a collaborative scheduling optimization module that uses the predicted load as the core constraint and solves the low-carbon, economical, and reliable generator output and power grid interaction scheme through a multi-objective optimization algorithm; an execution and monitoring module that converts the scheme into equipment instructions, collects operational data in real time, and corrects deviations; and an effect evaluation module that compares the predicted, planned, and actual values, outputs optimization suggestions, and feeds them back to the front-end module to form a closed loop. The entire system achieves full-process control from accurate load forecasting to optimized energy scheduling, real-time operation management, and continuous performance improvement, solving the problems of high carbon emissions, low efficiency, and unreliability in traditional industrial park energy scheduling.

[0039] Module 1: Multi-source data acquisition module

[0040] As the data source of the system, it accurately collects data from the entire agricultural machinery park, providing raw data support with granular matching and complete dimensions for subsequent modules, thus solving the prediction bias problem caused by the lag and lack of dimensions in traditional data collection.

[0041] Technical Implementation Details

[0042] 1. Data Collection Dimensions and Granularity Design

[0043] Load-related data: The data collection frequency is set differently according to the granularity of short-term (15 minutes), medium-term (1 hour), and long-term (1 day): agricultural machinery operation data (seeders, harvesters, etc.) are collected every 15 seconds (instantaneous power, operating status) and summarized into statistical values ​​for 15 minutes / 1 hour / 1 day; office and lighting power consumption is collected every minute and summarized into granularity for 15 minutes; historical load data is traced back to the past year and archived at the same granularity.

[0044] Agricultural machinery operation plan data: Collect agricultural machinery type (all categories such as seeders / harvesters / dryers), planned operating time (accurate to the minute), and operating power range (e.g., harvester 800-1000kW).

[0045] Environmental and power grid data: Temperature, humidity, and rainfall probability are collected every 15 minutes (for short-term forecasting), and daily average environmental data are collected (for long-term forecasting); peak-valley-flat electricity prices and carbon emission factors are updated daily (data cycle of synchronization module 2).

[0046] Unit operation data: The output, methanol consumption, and start / stop status of the methanol generator set are collected every 15 seconds to provide real-time data for modeling in Module 4.

[0047] The multi-source data acquisition module is also used to import methanol life cycle carbon emission factors from industry standard databases and obtain grid peak-valley-flat differential carbon emission factors from power company interfaces. The methanol life cycle carbon emission factors, grid carbon emission factors, and the collected real-time data together constitute a structured raw dataset. The subsequent data preprocessing module classifies the above carbon emission factors into carbon emission data in the standardized dataset, providing a basis for carbon emission calculation for the methanol generator set operation modeling module.

[0048] 2. Data Acquisition Equipment and Transmission Scheme

[0049] Equipment selection: Edge sensors (supporting RS485 protocol, accuracy ±0.5%) are deployed on agricultural machinery, smart meters (supporting NB-IoT, error ≤0.2) are used on office / lighting terminals, environmental data is obtained from the park's weather station (temperature accuracy ±0.1℃), and power grid data is obtained through the power system API interface.

[0050] Transmission architecture: Edge device data is aggregated via a 5G industrial gateway. Real-time data (every 15 seconds) is transmitted via UDP protocol, while historical summary data (15 minutes or more) is transmitted via TCP protocol.

[0051] 3. Preliminary data quality control

[0052] Deploy a local preprocessing unit: perform range verification (e.g., agricultural machinery power does not exceed the rated value ±20%) and abrupt change detection (power fluctuation >30% within 15 seconds triggers an alarm), remove obvious outliers before transmitting to module 2, reducing the pressure of subsequent processing.

[0053] Input and Output

[0054] Input: Agricultural machinery operation plan (imported from the park management system), power grid basic parameters (interface with the power company);

[0055] Output: The structured raw dataset (categorized by load association, work schedule, environment, power grid, and generating unit, including multi-granularity data of 15 seconds / 1 minute / 15 minutes / 1 hour / 1 day) is directly transferred to Module 2.

[0056] Module 2: Data Preprocessing Module

[0057] The original data in Module 1 is cleaned, normalized, and feature extracted to output a standardized dataset, which solves the problem of poor generalization ability of the prediction model caused by data heterogeneity and feature redundancy.

[0058] Technical Implementation Details

[0059] 1. Data Cleaning

[0060] Outlier handling: Combining the time series nature of load data, a dual filtering method of 3σ criterion + isolated forest algorithm is adopted. Short-term 15-minute granular data uses 3σ criterion (to remove data that deviates from the mean by 3 times the standard deviation), while long-term daily data uses isolated forest (to detect nonlinear anomalies, such as load mutations caused by extreme weather). Outliers are replaced with the weighted average of the three adjacent periods (with a weight of 0.6 for recent data).

[0061] Missing value imputation: For high-frequency data missing at 15 seconds / 1 minute, use linear interpolation (to ensure short-term continuity and adapt to ARIMA stability requirements); for data missing at granularity of 15 minutes and above, use replacement with the same day and trend correction (e.g., if the load is missing at 10:00 on a certain workday, use the average of the data from the same period of the last 3 workdays combined with the load trend of the day for adjustment).

[0062] Data alignment: All data are aligned uniformly according to the prediction time index t (t=1, 2, ..., T) to ensure that the 15-minute granular load data corresponds one-to-one with the agricultural machinery operation status and environmental data, and to avoid feature-label misalignment during training.

[0063] 2. Normalization processing

[0064] Numerical features (power, temperature, electricity price, etc.) are normalized to [0, 1] using Min-Max, as shown in the formula. ,in We take the extreme values ​​of the data from the past 30 days (to avoid the influence of historical outliers) to adapt to the stability requirements of XGBoost feature weight calculation and LSTM gradient descent.

[0065] Category-based features (agricultural machinery operating status, weather type) use unique hot coding, such as the code for a seeder in operation being [1, 0, 0] and the code for a machine stopped being [0, 1, 0], to ensure that they can be directly accessed.

[0066] 3. Feature Extraction

[0067] Short-term features (for ARIMA): Extract the load values ​​of the first three periods at a 15-minute granularity, the first-order difference of the load, and the number of short-term start-stop times of agricultural machinery, which directly support the calculation of the autoregressive term of the ARIMA(3,1,2) model.

[0068] Mid-term features (for XGBoost): Extract ARIMA predictions and residuals at the 1-hour granularity, the number of agricultural machines operating at the hourly level, the average ambient temperature, and the historical load for the same period, corresponding to the 12-dimensional feature vector of XGBoost.

[0069] Long-term features (for LSTM): Extract the XGBoost predictions and residuals, ARIMA predictions and residuals, daily average temperature, daily operating hours of agricultural machinery, and the load of the same period 7 days ago at the 1-day granularity, corresponding to the 15-dimensional feature vector of LSTM.

[0070] Feature selection: Redundant features are removed using the Pearson correlation coefficient (|r|>0.3). For example, features with low correlation between park humidity and load are deleted to ensure the efficiency of algorithm training.

[0071] Input and Output

[0072] Input: The structured raw dataset output by module 1;

[0073] Output: Standardized dataset (categorized by load type, operation type, environment type, power grid type, and unit type, including short-term / medium-term / long-term feature subsets).

[0074] Module 3: Electricity Load Forecasting Module for Agricultural Machinery Parks

[0075] I. Module Location and Algorithm Division of Labor

[0076] The core task of the electricity load forecasting module in the agricultural machinery park is to output high-precision load data covering the entire time period from 15 minutes to 30 days, providing a benchmark for subsequent methanol power generation and electricity consumption coordinated scheduling. Based on the load time-series characteristics, the module employs a three-level algorithm division of labor: the ARIMA algorithm focuses on short-term forecasting at a 15-minute granularity, accurately capturing stable time-series fluctuations; the XGBoost algorithm is responsible for medium-term forecasting at a 1-hour granularity, fusing multiple features to solve non-stationary problems; and the LSTM algorithm undertakes long-term forecasting at a 1-day granularity, capturing nonlinearity and long-period dependencies. The three algorithms interact pairwise through forward feature transfer and reverse residual correction, ultimately outputting a unified forecast result through weighted fusion, ensuring forecast accuracy across all time periods.

[0077] II. Core Algorithm Model Construction, Training, and Interaction

[0078] 1. ARIMA algorithm (ARIMA(3,1,2)): Short-term load forecasting (15-minute granularity)

[0079] The core value of the ARIMA algorithm lies in processing short-term stationary time-series data—the 15-minute granular load in the agricultural machinery park is mainly affected by stationary factors such as the single start-up and shutdown of the seeder and the short-term operation of office equipment, without significant nonlinear disturbances. After optimization by the AIC criterion (Akaike Information Criterion), the model order was determined to be ARIMA(3, 1, 2), namely, 3rd order autoregression (AR(3)), 1st order difference (I(1)), and 2nd order moving average (MA(2)).

[0080] 1.1 Model Construction

[0081] Model construction follows the logic of stationarization → time series fitting → inverse stationarization:

[0082] Step 1: First-order difference (to eliminate non-stationarity)

[0083] Original load sequence (t is a time index, dimensionless) T is the total timing length; The actual load at time t (in kW) is non-stationary due to the influence of agricultural machinery operation periods (ADF test: statistic = -1.8 < critical value -2.89), and needs to be stationary through first-order difference gradation. The formula is:

[0084] ;

[0085] in The first-order difference result of the load at time t (unit: kW, trend eliminated by differencing). After differencing, the ADF statistic = -4.2 > the critical value -2.89, confirming the stationarity of the sequence.

[0086] Step 2: AR(3) Autoregressive Term (Capturing Short-Term Dependencies)

[0087] Based on stationary sequences A third-order autoregressive term is constructed. The core logic is that the current difference result depends on the values ​​of the previous three time steps. The formula is as follows:

[0088] ;

[0089] in The AR(3) coefficients are dimensionless and obtained through maximum likelihood estimation. =0.62, =0.18, =0.09), Output the autoregressive term (unit: kW).

[0090] Step 3: MA(2) moving average term (corrected residuals)

[0091] A second-order moving average term is introduced to correct for random residuals (such as load fluctuations caused by sudden agricultural machinery failures). The formula is as follows:

[0092] ;

[0093] in The coefficients of MA(2) (dimensionless, estimated as follows) =-0.35, =-0.12), The residual at time t-1 (unit: kW) , (for predicted values) Output the moving average (unit: kW).

[0094] Step 4: Complete Model and Inverse Difference Reconstruction

[0095] Integrating AR(3) and MA(2), we obtain the complete ARIMA(3,1,2) model:

[0096] ;

[0097] in White noise sequence ( , =12.8, unit The original load forecast value is restored through inverse differential calculation:

[0098] ;

[0099] in This is the ARIMA algorithm's predicted value at time t (unit: kW, superscript (A) specifically refers to the ARIMA output). This represents the predicted value of the difference sequence (unit: kW).

[0100] 1.2 Model Training

[0101] The training process not only optimizes its own accuracy but also embeds cross-algorithm interactions:

[0102] 1. Data input: 15-minute granular normalized load sequence output from module 2 (672 samples in total over 7 days);

[0103] 2. Parameter optimization: After estimating the initial parameters using maximum likelihood estimation, XGBoost residuals are introduced. (The residual at time t of XGBoost, in kW) The inverse correction AR coefficient is calculated using the following formula:

[0104] ;

[0105] in =0.01 (learning rate, dimensionless). The gradient of the coefficients (calculated) ), where m is the number of iterations;

[0106] 3. Accuracy verification: Mean absolute percentage error of the corrected validation set. ( The mean absolute percentage error of the ARIMA algorithm, in %, is expressed by the formula: This meets the short-term accuracy requirements.

[0107] 1.3 Model Application

[0108] The application phase is the core of ARIMA's interaction with other algorithms:

[0109] 1. Predicted Output: Input short-term features from module 2 (15-minute granular historical load, short-term agricultural machinery operation plan), and output the model. ;

[0110] 2. Positive interaction: and (The ARIMA time-t residual, in kW) is passed to XGBoost as its input feature to help XGBoost correct mid-term prediction bias;

[0111] 3. Reverse interaction: Receiving feedback from XGBoost Update using the coefficient correction formula in step 2 Recalculate Obtain the correction value This further reduces errors.

[0112] 2. XGBoost Algorithm: Medium-Term Load Forecasting (1-Hour Granularity)

[0113] As an ensemble learning model, XGBoost's core advantage lies in its ability to integrate multiple features to handle medium-term non-stationary loads. One-hour load granularity is affected by multiple factors such as the start-up and shutdown of harvesters, changes in ambient temperature, and adjustments in grid electricity prices, requiring multi-feature fitting to achieve accurate prediction. Simultaneously, XGBoost acts as a bridge between short-term and long-term predictions, connecting prediction results at different granularities through bidirectional interaction with ARIMA and LSTM.

[0114] 2.1 Model Construction

[0115] The model, centered on multiple features and cross-algorithm interaction terms, ensures that medium-term predictions maintain short-term accuracy.

[0116] Step 1: Input feature vector definition

[0117] Feature vector (Units not specified, normalized) A total of 12 dimensions, covering all elements including agricultural machinery, environment, and short-term interactions. The formula is:

[0118] ;

[0119] : Percentage of agricultural machinery operation load (dimensionless, e.g., harvester load accounts for 40% of total load);

[0120] Park temperature (°C, hourly average temperature output by module 2);

[0121] Rainfall probability (%, hourly data output by module 2);

[0122] Number of seeders in operation (units, from agricultural machinery operation plan);

[0123] Number of harvesters in operation (units, same as above);

[0124] : Number of dryers in operation (units, same as above);

[0125] Number of irrigation pumps in operation (units, same as above);

[0126] Office power consumption (kW, hourly load output from module 2);

[0127] Lighting power consumption (kW, hourly load output from module 2);

[0128] ARIMA Predictions (kW, positive interaction term, supporting short-term accuracy).

[0129] ARIMA residuals (kW, positive interaction term, using short-term residuals to correct for medium-term bias);

[0130] Historical load for the same period (kW, such as the 1-hour load for the same period last week).

[0131] Step 2: Objective Function Design

[0132] The objective function includes a loss term (prediction bias) and a regularization term (to prevent overfitting), and also incorporates ARIMA residuals. The formula for calculating the correction deviation is:

[0133] ;

[0134] in The value predicted by the XGBoost algorithm at time t (unit: kW, with the superscript (X) specifically referring to the XGBoost output). =0.4 (ARIMA residual weight, dimensionless), K=100 (total number of decision trees). =0 (leaf node penalty coefficient) =1 (weight decay coefficient), =32 (the number of leaf nodes in the k-th tree). (Weight of the leaf node of the k-th tree, in kW).

[0135] Step 3: Iterative Updates and Final Prediction

[0136] In each iteration, ARIMA information is incorporated when generating a new decision tree. The formula is as follows:

[0137] ;

[0138] Where m is the iteration round. This represents the output (in kW) of the k-th tree. For decision tree sets, This is the regularization term for a single tree. By integrating the outputs of all trees and incorporating ARIMA prediction corrections, the final prediction formula is obtained:

[0139] ;

[0140] in =0.2 (ARIMA prediction weight, dimensionless).

[0141] 2.2 Model Training

[0142] The training process focuses on multi-feature fitting and cross-algorithm optimization:

[0143] 1. Data Input and Splitting: Input module 2's hourly dataset (720 samples over 30 days) + ARIMA output. and The dataset was divided into a training set (576 samples) and a validation set (144 samples) in an 8:2 ratio.

[0144] 2. Hyperparameter optimization: The optimal parameters are obtained through grid search: learning rate. =0.1, tree depth =6. Number of iterations =100;

[0145] 3. LSTM Interactive Correction: Introducing LSTM Residuals (The residual at time t of the LSTM, in kW) is used to update the decision tree weights, using the following formula:

[0146] ;

[0147] in =0.001 (learning rate, dimensionless). =0.85 (weight gradient);

[0148] 4. Accuracy verification: Corrected validation set ( The mean absolute percentage error of XGBoost (in %) meets the mid-term accuracy requirements.

[0149] 2.3 Model Application

[0150] In the application phase, reverse correction with ARIMA and forward output with LSTM are implemented:

[0151] 1. Predictive Execution: Input module 2's mid-term features (hourly agricultural machinery operation plans, environmental data) + ARIMA's... and Output ;

[0152] 2. Positive interaction: and (The residual at time t of XGBoost, in kW) is passed to LSTM as its core input feature to provide medium-term trend information for long-term prediction;

[0153] 3. Reverse interaction: Receiving ARIMA's... With LSTM The decision tree parameters are adjusted using the weight update formula in step 2 to correct the error. get .

[0154] 3. LSTM Algorithm: Long-Term Load Forecasting (1-Day Granularity)

[0155] As a recurrent neural network, LSTM's core advantage lies in capturing the nonlinearity and long-term temporal dependencies of long-term loads through gating mechanisms. Daily load granularity is influenced by factors such as agricultural machinery operation cycles (e.g., high load throughout the day during the sowing season), weekly patterns (e.g., reduced operation on weekends), and environmental trends (e.g., increased irrigation load due to continuous high temperatures), requiring the processing of cross-day temporal correlations. LSTM compensates for its own short-term accuracy limitations through interaction with ARIMA / XGBoost.

[0156] 3.1 Model Construction

[0157] The model uses a gating mechanism and cross-algorithm interaction features as its core to ensure that long-term predictions closely match short- and medium-term trends.

[0158] Step 1: Network Structure Design

[0159] Constructing a 3-layer LSTM network:

[0160] Input layer: 15 dimensions (corresponding to feature vectors) ), activation function ReLU;

[0161] Hidden layers: 2 layers, 128 neurons per layer, activation function tanh, dropout rate 0.2;

[0162] Output layer: Dimension 1 (corresponding to) ), activation function Linear;

[0163] Optimizer: Adam, Learning Rate =0.001, =0.9, =0.999.

[0164] Step 2: Input feature vector definition

[0165] Feature vector (Unitless, normalized) A total of 15 dimensions, the formula is:

[0166] ;

[0167] : XGBoost predicted value / residual (kW, positive interaction term);

[0168] : ARIMA predicted value / residual (kW, positive interaction term);

[0169] : Daily average temperature (°C, daily environmental data output by module 2);

[0170] : Daily rainfall probability (%, daily environmental data output by module 2);

[0171] Total daily agricultural machinery operation time (h, from the operation plan);

[0172] : Number of seeders, harvesters, dryers, and irrigation pumps in operation per day (units, from work plan);

[0173] Peak, flat, and valley electricity prices (RMB / kWh, grid data output from module 2);

[0174] : Daily load (kW) 7 days ago (capturing weekly periodicity).

[0175] Step 3: Core formula of gating mechanism (time series information processing)

[0176] LSTM achieves selective storage and updating of time-series information through input gates, forget gates, and output gates. The core formula is based on the input at time t. Hidden layer output at time t-1 (Unit: kW), Cell state (Unit: kW, Core Memory Unit):

[0177] 1. Input Gate (controls the entry of new information):

[0178] ;

[0179] in The input gate outputs a dimensionless value [0, 1]. (for the Sigmoid function). (128×15) (128×128) is the weight matrix (unit: 128×128). ), (128×1) is the bias vector (unit: kW).

[0180] 2. Forget Gate (Controls the retention of historical information):

[0181] ;

[0182] in Output the forget gate (dimensionless, value [0, 1]). Weight matrix (unit) ), This is the bias vector (unit: kW).

[0183] 3. Cell state update:

[0184] ;

[0185] ;

[0186] in Candidate cell status (unit: kW). The updated cell state (unit: kW). This is the Hadamard product (element-wise multiplication). Weight matrix (unit) ), This is the bias vector (unit: kW).

[0187] 4. Output gate (control information output):

[0188] ;

[0189] ;

[0190] in The output gate outputs a dimensionless value [0, 1]. Output of the hidden layer (unit: kW). Weight matrix (unit) ), This is the bias vector (unit: kW).

[0191] Step 4: Final Prediction Formula

[0192] The output layer integrates the hidden layer output and cross-algorithm interaction terms, as shown in the formula:

[0193] ;

[0194] in The value predicted by the LSTM algorithm at time t (unit: kW, superscript (L) specifically refers to the LSTM output). (1×128) represents the output layer weights (units). ), For bias (unit: kW). =0.3 (XGBoost weights, dimensionless). =0.1 (ARIMA weight, dimensionless).

[0195] 3.2 Model Training

[0196] The training process revolves around long-term fitting and cross-algorithm optimization:

[0197] 1. Data Input and Preprocessing: Input module 2's daily dataset (90 samples over 90 days) + ARIMA / XGBoost output , The features are mapped to [0, 1] by Min-Max normalization;

[0198] 2. Network training: using mean squared error (MSE) as the loss function ( (50 iterations, with backpropagation updating the weights in each iteration);

[0199] 3. Cross-algorithm interactive correction: LSTM residuals Feedback is sent to ARIMA / XGBoost, and the corrected predictions from both are received simultaneously to adjust the weights of its own features.

[0200] 4. Accuracy verification: Validation set =5.8% ( The mean absolute percentage error of the LSTM (in %) satisfies long-term accuracy requirements.

[0201] 3.3 Model Application

[0202] LSTM primarily provides optimization data for ARIMA / XGBoost through reverse interaction:

[0203] 1. Predictive Execution: Input long-term features from module 2 (daily agricultural machinery operation plan, environmental prediction data) + ARIMA / XGBoost predicted values / residuals, output... ;

[0204] 2. Reverse interaction: This information is passed to ARIMA / XGBoost as the basis for parameter updates in both. ARIMA then... Adjustment XGBoost via Update the decision tree weights to achieve global prediction optimization.

[0205] III. Fusion Processing of Three Algorithms

[0206] The three algorithms achieve global optimization through pairwise bidirectional interaction and error weighting.

[0207] 1. Pairwise interaction process

[0208] (1) Interaction between ARIMA and XGBoost

[0209] Positive Interaction: ARIMA and As input features of XGBoost ( XGBoost via =0.4 and A weight of 0.2 is incorporated into the forecast to reduce mid-term bias;

[0210] Reverse interaction: XGBoost Feedback is sent to ARIMA and corrected using the coefficient formula. Update ARIMA Improve short-term accuracy.

[0211] (2) Interaction between XGBoost and LSTM

[0212] Positive interaction: XGBoost and As the core input feature of LSTM, LSTM uses... A weight of 0.3 incorporates medium-term trends into long-term forecasts;

[0213] Reverse interaction: LSTM Feedback is sent to XGBoost and updated via the weight update formula. Adjust the decision tree weights to optimize medium-term forecasts.

[0214] (3) Interaction between ARIMA and LSTM

[0215] Positive Interaction: ARIMA and As input features for LSTM, LSTM uses... A weight of 0.1 is used to supplement short-term correction information;

[0216] Reverse interaction: LSTM Feedback is sent to ARIMA to help adjust the AR coefficient and further optimize the stability of short-term forecasts.

[0217] 2. Overall Integration Formula and Calculation Process

[0218] The fusion process is carried out in four steps:

[0219] Step 1: Calculate the accuracy index (MAPE) for each algorithm.

[0220] Based on the actual load of module 2 Calculate the three algorithms :

[0221] ;

[0222] ;

[0223] ;

[0224] in This is the predicted value after reverse interaction correction.

[0225] Step 2: Assign initial weights using the inverse error ratio method

[0226] The initial weights are positively correlated with the algorithm's accuracy (the smaller the error, the larger the weight), as shown in the formula:

[0227] ;

[0228] ;

[0229] ;

[0230] in The initial weights (dimensionless) for the three algorithms.

[0231] Step 3: Introduce interaction coefficients to correct weights

[0232] Interaction coefficients are determined based on pairwise interaction strength. (Dimensionless) as follows:

[0233] (The strength of the interaction between ARIMA and XGBoost); (The interaction strength between ARIMA and LSTM);

[0234] (XGBoost interaction strength with ARIMA); (The interaction strength of XGBoost with LSTM);

[0235] (The interaction strength of LSTM with ARIMA); (The interaction strength of LSTM with XGBoost).

[0236] The final weight formula after correction is:

[0237] ;

[0238] ;

[0239] ;

[0240] like Then normalization is performed. (i = A, X, L), ensuring the weight sum is 1.

[0241] The Pearson correlation coefficient between the predicted values ​​and residuals of different algorithms is calculated to quantify the degree of linear correlation between the algorithms. The magnitude of the correlation coefficient directly determines the initial interval of the interaction coefficient. Then, multiple rounds of iterative verification are performed using historical datasets, and the coefficient values ​​are adjusted according to the changes in the accuracy of the fusion prediction.

[0242] Step 4: Final Fusion Prediction Formula

[0243] Integrate and correct the predictions from the three algorithms to output a unified prediction result:

[0244] ;

[0245] in This is the final load forecast at time t after fusion (unit: kW, superscript (F) specifically refers to the fusion result).

[0246] 3. Example of fused computing

[0247] Taking t=10 (corresponding to 2 hours and 30 minutes, 15-minute granularity) as an example, the complete fusion process is demonstrated:

[0248] 1. Basic data: Actual load =850kW; ARIMA correction =848.5kW, =3.2%; XGBoost correction =855kW =4.2%; =840kW =5.8%.

[0249] 2. Calculate the initial weights:

[0250] =0.3125, ≈0.2381, ≈0.1724;

[0251] =0.3125 / (0.3125+0.2381+0.1724)≈0.432;

[0252] ≈0.329, ≈0.239.

[0253] 3. Calculate the corrected weights:

[0254] =0.432+0.1×0.329+0.05×0.239=0.47685;

[0255] =0.329+0.12×0.432+0.08×0.239=0.400;

[0256] =0.239+0.06×0.432+0.09×0.329=0.29453;

[0257] After normalization: ≈0.407, ≈0.341, ≈0.252.

[0258] 4. Final fusion result:

[0259] =0.407×848.5+0.341×855+0.252×840≈848.58kW;

[0260] After fusion =0.16%, far lower than the accuracy of a single algorithm.

[0261] IV. Core Roles and Contributions of Each Algorithm

[0262] 1. ARIMA Algorithm: Its core function is to provide a short-term, high-precision benchmark by capturing the stationary time-series characteristics of a 15-minute granular load and outputting... As the basic input of XGBoost / LSTM, the fusion contribution improves the short-term fit of the full-time prediction by forward propagation of the predicted value / residual and backward reception of correction information, thus avoiding high-frequency deviations in the medium-term prediction.

[0263] 2. XGBoost Algorithm: Its core function is to bridge short-term and long-term forecasts. It solves the non-stationarity of 1-hour granularity load by fusing multiple features, and at the same time, it transfers short-term accuracy to long-term forecasts. The fusion contribution acts as an intermediate hub, transforming short-term information from ARIMA into medium-term trends and then passing it to LSTM. At the same time, it receives long-term feedback from LSTM to optimize itself, achieving a seamless connection between short-term, medium-term, and long-term forecasts.

[0264] 3. LSTM Algorithm: Its core function is to capture long-term nonlinearity and long-period dependence, outputting 1-day granular load forecasts; the fusion contribution is achieved by backpropagating the residuals. It corrects the parameters of ARIMA / XGBoost to avoid them getting stuck in local optima, while receiving short- and medium-term information to make up for its own short-term lack of accuracy, ensuring that long-term forecasts are both in line with trends and do not lose details.

[0265] 4. Fusion Algorithm: The process of interactive fusion of the three load forecasting algorithms is as follows: Figure 3As shown, its core function is to integrate the advantages of the three algorithms and eliminate the limitations of a single algorithm (ARIMA cannot handle nonlinearity, XGBoost has weak long-period fitting, and LSTM has low short-term accuracy); the contribution of the fusion is to output high-precision prediction results covering the entire time period from 15 minutes to 30 days. The time-segmented load forecast result set includes short-term forecasts at a 15-minute granularity, medium-term forecasts at a 1-hour granularity, long-term forecasts at a 1-day granularity, and corresponding error rates, providing reliable and continuous load benchmark data for the methanol power generation-consumption coordinated scheduling optimization of Module 5.

[0266] It is important to note that the completion of the algorithm interaction loop in Module 3 requires meeting three conditions: "internal iterative convergence + cross-algorithm deviation compliance + full system evaluation compliance". Specifically:

[0267] Algorithm internal iterative convergence conditions

[0268] Each algorithm's own parameter iteration (such as the coefficient iteration of ARIMA, the decision tree weight iteration of XGBoost, and the network weight iteration of LSTM) reaches a convergence state, that is, in multiple consecutive iterations, the error of the algorithm's predicted value (such as the mean absolute percentage error) changes by less than a set threshold, and the parameter update value tends to be stable without significant fluctuations.

[0269] Cross-algorithm interaction deviation criteria

[0270] After pairwise reverse feedback correction, the reduction in prediction error caused by residual propagation between any two algorithms reaches a stable level: that is, after ARIMA receives residual correction from XGBoost / LSTM, its short-term prediction error no longer decreases significantly; after XGBoost receives residual correction from ARIMA / LSTM, its medium-term prediction error no longer decreases significantly; and after LSTM receives prediction values / residuals from ARIMA / XGBoost, its long-term prediction error no longer decreases significantly.

[0271] System-wide assessment of compliance conditions

[0272] The prediction results after the fusion of the three algorithms were evaluated by Module 7. The "prediction fit" (correlation coefficient between the actual load and the fused prediction value) reached the preset target, and the fit remained stable over multiple consecutive evaluation periods. At the same time, the prediction deviation had no significant negative impact on the economic, low-carbon, and reliability objectives of the collaborative scheduling scheme of Module 5. In other words, the deviation between the actual execution effect of the scheduling scheme and the planned value was controlled within an acceptable range.

[0273] When all three conditions are met, the algorithm interaction loop of module 3 completes one full iteration; if not, the algorithm parameter correction and interaction are continuously triggered until all termination conditions are met.

[0274] Module 4: Methanol Generator Set Operation Modeling Module

[0275] A quantitative model and operational constraints for methanol generator set output, energy consumption, cost, and carbon emissions are established to provide load matching unit parameters for module 5 (cooperative scheduling). Based on the full-time load predicted by module 3, the operating characteristics of the unit under different outputs are accurately output to ensure that the scheduling scheme meets both load demand and unit physical limitations.

[0276] Technical Implementation Details

[0277] 1. Output-Energy Consumption Model (15-minute granularity, prediction accuracy of matching module 3)

[0278] Based on the unit operation data (output P, ​​methanol consumption Q) output by Module 2, a piecewise linear fitting model is constructed: when P≤500kW, Q=0.8P+20; when 500<P≤1000kW, Q=0.75P+45; when P>1000kW, Q=0.9P (reflecting the unit's characteristic of low efficiency at low load and high efficiency at rated load), and the fitting error is controlled within ±2%.

[0279] Model validation: Testing with 15-minute granular data from the past 7 days ensures that the predicted energy consumption values ​​for different output ranges deviate from the actual values ​​by ≤3%, providing an accurate basis for calculating methanol costs in Module 5.

[0280] 2. Cost and Carbon Emission Model (Time-Specific Quantification)

[0281] Cost model: Total operating cost = methanol cost + maintenance cost + start-up and shutdown cost, where methanol cost = energy consumption Q × methanol unit price (updated daily), maintenance cost = output P × 0.02 yuan / kWh (based on unit operation and maintenance manual), start-up and shutdown cost = 1500 yuan / time (including methanol loss and mechanical wear), and the cost calculation granularity is consistent with the prediction granularity of module 3 (15 minutes).

[0282] Carbon emission model: Total carbon emissions = Generating unit carbon emissions + Grid electricity purchase carbon emissions, where generating unit carbon emissions = energy consumption Q × 0.65 kg / L (methanol carbon emission factor, referencing industry standards), grid carbon emissions = purchased electricity × grid carbon emission factor (provided by module 2, with peak-valley-flat differentiation), directly supporting the low-carbon target optimization of module 5.

[0283] 3. Runtime constraint modeling (to ensure scheduling feasibility)

[0284] Output constraints: The rated output of the unit is 1200kW, and the minimum stable output is 200kW, that is, 200≤P≤1200kW (15-minute granularity), to avoid output exceeding the limit when Module 5 is scheduled.

[0285] Start-up and shutdown constraints: Each start-up and shutdown interval is ≥30 minutes, the start-up time is ≤5 minutes (the time from shutdown to rated output), and the shutdown time is ≤2 minutes. This constrains the unit start-up and shutdown plan of module 5.

[0286] Ramp-up constraint: Output change rate ≤ 100kW / minute (increasing or decreasing). If the output increases from 500kW to 800kW within 15 minutes, it needs to be adjusted in 3 stages (each time +100kW) to ensure that the dispatching plan is executable.

[0287] Reliability constraints: If the unit operates continuously for ≤24 hours, it needs to be shut down for maintenance for ≥1 hour to avoid unit failure caused by excessive scheduling of module 5.

[0288] Input and Output

[0289] Inputs: Standardized datasets output from Module 2 (unit-type data: historical output, methanol consumption, maintenance records), and time-of-use load forecast results set output from Module 3;

[0290] Output: Methanol generator set operation model set (including output-energy consumption model, cost model, carbon emission model, and operation constraints, which are directly transmitted to module 5 as optimization constraints).

[0291] Module 5: Cooperative Scheduling Optimization Module

[0292] Using the time-of-use load forecast result set output by Module 3 as the core constraint, combined with the unit model of Module 4 and the power grid data of Module 2, a multi-objective optimization algorithm is used to solve the optimal dispatch scheme of low carbon, economy and reliability, so as to solve the problems of poor load matching and single objective in traditional dispatch.

[0293] Technical Implementation Details

[0294] 1. Construction of the objective function

[0295] Low-carbon target: Total carbon emissions ,in (Unit carbon emissions) = Module 4 carbon emissions model × Unit output , (Grid carbon emissions) = Grid electricity purchases ×Grid carbon emission factor (provided by Module 2). ( (Integrate predicted loads for module 3 to ensure load requirements are met).

[0296] Economic Objective: Total Cost The various costs are calculated using the cost model in Module 4. (Grid Cost) = Purchased Electricity × Peak-Valley Price - Sold Electricity × Grid-Connected Price (at the time of sale) (Negative).

[0297] Reliability target: Load fulfillment rate The load is directly associated with module 3 to predict the load and ensure reliable power supply.

[0298] 2. Implementation of multi-objective optimization algorithm (MOPSO + entropy-weighted TOPSIS)

[0299] Initialization: Particle dimension = 3 × T (T is the total number of granular steps in 15 minutes, e.g., T = 96 in 1 day), including (Unit output) (Electricity purchase / sales) (Start-stop state, 1 = start, 0 = stop), number of particles 50, number of iterations 100, inertia weight 0.7 (balancing global search and local convergence).

[0300] Fitness function: Low-carbon and economic goals are normalized and weighted (initial weights are both 0.4), with the reliability goal as a constraint (if not satisfied, the fitness is 0). The formula is as follows: ,in This represents the maximum value under extreme operating conditions.

[0301] Constraint handling: The unit constraints (output, start-up, and ramping) and grid constraints (purchased electricity ≤ 500kW, sold electricity ≤ 300kW) in embedded module 4 are eliminated using the penalty function method (if the output exceeds the limit, the fitness is reduced by 0.5).

[0302] Solution decision: The Pareto optimal solutions are ranked using the entropy-weighted TOPSIS method, with the weights dynamically adjusted according to the feedback from module 7 (initial low carbon 0.4, economy 0.4, reliability 0.2), and the solution with the highest comprehensive score is selected.

[0303] Input and Output

[0304] Inputs: The methanol generator set operation model set output from module 4, and the time-of-use load prediction result set output from module 3. ), and the standardized datasets output by Module 2 (unit-type / cost-type / carbon emission-type data);

[0305] Output: The optimal collaborative scheduling scheme set, including methanol generator output plan, power grid purchase and sale plan, and generator start-up and shutdown plan with a 15-minute granularity.

[0306] Module 6: Scheduling Instruction Execution and Real-time Monitoring Module

[0307] The optimal scheduling scheme of Module 5 is transformed into executable instructions between the generator and the power grid. Through real-time data acquisition and deviation correction, the consistency between actual operation and plan is ensured. The core correlation between the predicted load of Module 3 and the planned value of Module 5 solves the deviation problem between prediction, scheduling and execution.

[0308] 1. Instruction conversion

[0309] Unit instruction: Move module 5 (15-minute particle size output) is converted into digital instructions, such as 10:00-10:15, methanol generator set output 800kW, speed 1500r / min, methanol supply 60L / min, instruction format adapted to the unit PLC control system (supports Modbus-RTU protocol).

[0310] Power grid instruction: (Purchased / sold electricity power) is converted into grid interaction commands, such as 10:00-10:15, the grid purchases 0kW of electricity and sales are prohibited, and the commands are transmitted to the park's distribution network control terminal through the power system interface.

[0311] Instruction verification: Compare with module 3 With Module 5 Ensure that the load fulfillment rate is ≥99.9%; otherwise, return to module 5 for re-optimization.

[0312] 2. Real-time data acquisition and deviation analysis

[0313] Data acquisition frequency: 15 seconds / acquisition of actual unit output Methanol consumption Electricity purchase / sale Actual load of the park (In conjunction with the data acquisition device in Module 1).

[0314] Deviation calculation: according to the formula Calculate the deviation, where The planned value for module 5. (The deviation between the actual load and the prediction of module 3), when or Corrections are triggered at specific times.

[0315] 3. Real-time correction and anomaly alarm

[0316] Correction Algorithm: The PI control algorithm is used to adjust the command, and the formula is as follows: ,in =0.8 (proportional coefficient, rapid response deviation). =0.2 (integral coefficient, to eliminate steady-state error). For example, if the actual output of the unit is 700kW (planned 800kW), then... The command was adjusted to output 780kW.

[0317] Abnormal alarm: When the deviation is continuously greater than 10% or the unit malfunctions (such as methanol leakage), a graded alarm is triggered: Level 1 alarm (SMS notification to maintenance personnel) and Level 2 alarm (automatic shutdown and switching to grid power supply). The alarm information includes time, type, actual value, planned value, and deviation value, such as 2025-11-28 10:05, the unit output deviation exceeds the limit (actual 700kW, planned 800kW, deviation -12.5%).

[0318] Input and Output

[0319] Inputs: The optimal coordinated scheduling scheme set output by module 5, and the time-sharing load forecast result set output by module 3;

[0320] Output: 1. Real-time operating status dataset (15-second granularity, including actual output, methanol consumption, electricity purchase / sale, load, deviation value, and correction instructions); 2. Anomaly alarm information (hierarchical classification, transmitted to module 7 and the park operation and maintenance system).

[0321] Module 7: Scheduling Performance Evaluation Module

[0322] With the predicted value of Module 3, the planned value of Module 5, and the actual value of Module 6 as the core, the scheduling effect is evaluated from four dimensions: economy, low carbon, reliability, and fit. Optimization suggestions are output and fed back to the front-end module, forming a closed loop of data collection, prediction, scheduling, evaluation, and optimization.

[0323] Technical Implementation Details

[0324] 1. Evaluation Dimensions and Quantification Methods

[0325] Economic assessment: Calculate the deviation rate between actual total cost and planned cost ( Unit electricity cost () ), cost savings rate ( , The historical no-schedule cost provided for Module 2 is used to evaluate the economics of the Module 5 scheme, while analyzing the impact of the prediction deviation of Module 3 on the cost (such as the over-predicted load leading to excessive unit output).

[0326] Low-carbon performance assessment: Calculate the deviation rate between actual total carbon emissions and planned carbon emissions, and carbon emissions per unit of electricity. ), carbon emission reduction rate ( , (For carbon emissions from pure grid power supply), the matching degree between the carbon emission model in module 4 and the load prediction in module 3.

[0327] Reliability assessment: Calculating power supply reliability ( ), load fulfillment rate ( ), unit start-up and shutdown success rate ( To ensure that the target of ≥99.9% is met, if it is not met, the prediction deviation of module 3 or the rationality of the constraint of module 4 will be analyzed.

[0328] Scheduling fit assessment: Computer group output fit (Pearson correlation coefficient between actual output and planned value) ), grid interaction fit (actual electricity purchase / sale versus planned value) ), Predictive fit (actual load and fused prediction value of module 3) (r>0.9 is excellent, r<0.7 needs optimization).

[0329] 2. Evaluation Report and Closed-Loop Feedback

[0330] Report content includes numerical values ​​for each dimension of indicators, analysis of the reasons for deviations (such as low prediction accuracy due to ARIMA failing to capture sudden shutdowns of agricultural machinery), and optimization suggestions (divided into modules). The format is standardized so that the system can automatically extract feedback information.

[0331] Feedback to Module 3: Adjust algorithm parameters based on predicted fit, such as... When <0.7, optimize the autoregressive coefficient of ARIMA. (Adjusted from 0.62 to 0.68), increased the number of hidden layer neurons in LSTM to 156, and adjusted the number of decision trees in XGBoost to 120.

[0332] Feedback to Module 1: Supplement the data sources with large acquisition deviations, such as the failure to collect the signal of sudden shutdown of agricultural machinery, which leads to prediction deviations, and add agricultural machinery fault sensors.

[0333] Feedback to Module 5: Adjust the weights of MOPSO / TOPSIS. If the low-carbon target is not met, increase the low-carbon weight from 0.4 to 0.5.

[0334] Input and Output

[0335] Inputs: Real-time operating status dataset and abnormal alarm information output by module 6; optimal scheduling scheme set (planned value) output by module 5; standardized dataset (historical no-scheduling data and pure grid carbon emission data) output by module 2; time-of-use load forecast result set (forecast value) output by module 3.

[0336] Output: Scheduling effect evaluation report (including closed-loop optimization suggestions, transmitted to modules 1, 3, and 5).

[0337] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0338] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0339] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0340] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0341] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0342] Furthermore, this application uses specific terms to describe embodiments of the application. For example, "an embodiment," "an embodiment," and / or "a number of embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of the application. Therefore, it should be emphasized and noted that an embodiment, an embodiment, or an alternative embodiment mentioned twice or more in different locations in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of the application can be appropriately combined.

[0343] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as units, modules, or systems. Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0344] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0345] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0346] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A low-carbon agricultural machinery park methanol power-electricity co-scheduling system, characterized in that, The application relates to a multi-source data acquisition module, a data preprocessing module, a load prediction module, a methanol generator set operation modeling module, a collaborative scheduling optimization module, a scheduling instruction execution and real-time monitoring module and a scheduling effect evaluation module. The multi-source data acquisition module is used for collecting load correlation data, agricultural machine operation plan data, environment data, power grid data and methanol generator set operation data of an agricultural machine park and outputting a structured original data set. The data preprocessing module is used for preprocessing the structured original data set, extracting short-term, medium-term and long-term three-level feature subsets and outputting a standardized data set. The load prediction module is used for receiving the standardized data set, adopting a self-regression integral moving average algorithm, an extreme gradient boosting algorithm and a long short-term memory network algorithm to construct a prediction model, taking short-term prediction values and residual errors output by the self-regression integral moving average algorithm as input features of the extreme gradient boosting algorithm, taking medium-term prediction values and residual errors output by the extreme gradient boosting algorithm as input features of the long short-term memory network algorithm, feeding residual errors output by the long short-term memory network algorithm back to the self-regression integral moving average algorithm and the extreme gradient boosting algorithm to correct parameters of the two algorithms, and outputting a time-period load prediction result set through weighted fusion of the three algorithms after interaction correction. The methanol generator set operation modeling module is used for receiving unit data of the standardized data set, constructing an output-energy consumption model, a cost model, a carbon emission model and an operation constraint condition and outputting a methanol generator set operation model set. The collaborative scheduling optimization module is used for receiving the time-period load prediction result set, the methanol generator set operation model set and power grid data of the standardized data set, solving an optimal scheme through a multi-objective particle swarm optimization algorithm combined with an entropy weight approximation ideal solution sorting method and outputting an optimal collaborative scheduling scheme set.

2. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 1, characterized in that, The scheduling instruction execution and real-time monitoring module is used for receiving the optimal collaborative scheduling scheme set and the time-period load prediction result set, converting the scheme into equipment execution instructions, collecting actual operation data in real time and calculating deviations from planned values, adjusting the instructions through a deviation correction algorithm, outputting a real-time operation state data set and abnormal alarm information and feeding back the optimal collaborative scheduling scheme set and the time-period load prediction result set. The scheduling effect evaluation module is used for receiving the real-time operation state data set, the abnormal alarm information, the optimal collaborative scheduling scheme set and historical benchmark data of the standardized data set, evaluating effects from the dimensions of economy, low carbon, reliability and scheduling fitting degree, outputting an evaluation report containing optimization suggestions and feeding back the evaluation report to the multi-source data acquisition module, the load prediction module and the collaborative scheduling optimization module. The weighted fusion process of the load prediction module is specifically as follows: average absolute percentage errors of the self-regression integral moving average algorithm, the extreme gradient boosting algorithm and the long short-term memory network algorithm are calculated, and initial weights of the three algorithms are distributed according to the error inverse rule.

3. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 1, characterized in that, ​ The initial weight is corrected by introducing an interaction coefficient between algorithms, including interaction intensity of the autoregressive integral moving average algorithm on the extreme gradient boosting algorithm, interaction intensity of the extreme gradient boosting algorithm on the autoregressive integral moving average algorithm, interaction intensity of the extreme gradient boosting algorithm on the long short-term memory network algorithm, and interaction intensity of the long short-term memory network algorithm on the autoregressive integral moving average algorithm and the extreme gradient boosting algorithm. The final weight after correction satisfies a total sum of 1. The fused prediction value is calculated by the corrected prediction value of the autoregressive integral moving average algorithm multiplied by the weight of the autoregressive integral moving average algorithm, the corrected prediction value of the extreme gradient boosting algorithm multiplied by the weight of the extreme gradient boosting algorithm, and the prediction value of the long short-term memory network algorithm multiplied by the weight of the long short-term memory network algorithm.

4. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 1, characterized in that, The preprocessing and feature extraction of the data preprocessing module specifically include: processing abnormal values by using 3σ criterion + isolated forest algorithm, 3σ criterion for 15-minute and below granularity data, and isolated forest algorithm for 1-hour and above granularity data, replacing abnormal values with weighted mean values of adjacent 3 periods, and recent data weight being 0.6; filling missing values by using linear interpolation or similar day replacement + trend correction; when extracting features, extracting 15-minute granularity previous 3-period load values, load first-order difference, and agricultural machinery short-time start-stop times for the autoregressive integral moving average algorithm, extracting 1-hour granularity prediction value of the autoregressive integral moving average algorithm, residual error of the autoregressive integral moving average algorithm, agricultural machinery hour-level running number, and historical same-period load for the extreme gradient boosting algorithm, and extracting 1-day granularity prediction value of the extreme gradient boosting algorithm, residual error of the extreme gradient boosting algorithm, prediction value of the autoregressive integral moving average algorithm, residual error of the autoregressive integral moving average algorithm, and 7-day previous same-period load for the long short-term memory network algorithm; and forming a three-level feature subset after the extracted features are screened by Pearson correlation coefficient.

5. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 1, characterized in that, The output-energy consumption model of the methanol generator set operation modeling module is a piecewise linear fitting model: taking 50% of the rated output of the methanol generator set as a dividing point, different fitting coefficients are set for the low-load section and the high-load section; the fitting data are derived from the historical data of the unit in the standardized data set, including historical output values and corresponding methanol consumption, and the model fitting error is controlled within ±2%.

6. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 1, characterized in that, The low-carbon, economic, and reliable targets of the collaborative scheduling optimization module are specifically: the low-carbon target is to minimize the sum of carbon emissions of the methanol generator set and carbon emissions of grid power purchase, the economic target is to minimize the sum of methanol consumption cost, maintenance cost, start-stop cost, and grid power purchase and sale cost, and the reliable target is to make the load satisfaction rate ≥99.9%; in the multi-objective particle swarm optimization algorithm, the particle dimension includes the output of the methanol generator set, the grid power purchase and sale power, and the unit start-stop state, the inertia weight is set to 0.7 to balance the global search and local convergence ability, and the constraint conditions include 200kW≤P≤1200kW, ramp rate ≤100kW / minute, start-stop interval ≥30 minutes, grid power purchase ≤500kW, and power sale ≤300kW.

7. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 2, characterized in that, The deviation correction algorithm of the scheduling instruction execution and real-time monitoring module is a PI control algorithm, and the proportional coefficient in the correction instruction calculation formula is 0.8 to quickly respond to the deviation, and the integral coefficient is 0.2 to eliminate the steady-state error; the condition for triggering correction is that the absolute value of the deviation between the actual output of the equipment and the planned value is greater than 5% of the planned value, or the absolute value of the deviation between the actual load of the park and the set of time period load prediction results is greater than 3% of the predicted value.

8. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 2, characterized in that, The optimization suggestions of the scheduling effect evaluation module specifically include: outputting prediction deviation analysis to the load prediction module for adjusting the autoregressive coefficient of the autoregressive integrated moving average algorithm, the number of decision trees of the extreme gradient boosting algorithm, and the number of hidden layer neurons of the long short-term memory network algorithm; outputting target achievement conditions to the collaborative scheduling optimization module for dynamically adjusting the weights of low-carbon, economic and reliable targets; and outputting deviation data source analysis to the multi-source data acquisition module for supplementing the acquisition of agricultural machinery failure signals, extreme environment parameters and power grid voltage fluctuation data.

9. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 1, characterized in that, The collection granularity of the multi-source data acquisition module adapts to the three-level prediction needs of the load prediction module: agricultural machinery operation data is collected at 15 seconds / time, office and lighting electricity is collected at 1 minute / time, environmental data is collected at 15 minutes / time, and power grid data and agricultural machinery operation plan data are updated daily; all collected data are summarized into structured original data sets of 15 minutes, 1 hour and 1 day granularity to ensure the granularity matching of short-term, medium-term and long-term feature subsets.

10. The low-carbon farm machinery park methanol power-generation-power consumption collaborative scheduling system according to claim 1, characterized in that, The carbon emission model of the methanol generator set operation modeling module adopts factor fusion calculation, and the total carbon emission amount = methanol generator set carbon emission amount + power grid purchased electricity carbon emission amount; wherein the methanol generator set carbon emission amount = methanol consumption amount x methanol full life cycle carbon emission factor, and the power grid purchased electricity carbon emission amount = power grid purchased electricity amount x power grid peak valley difference differentiated carbon emission factor, both of which are derived from the carbon emission data in the standardized data set.

Images