Method and device for predicting short-term load of integrated energy system for agricultural park

Abstract

An agricultural park-oriented comprehensive energy system short-term load prediction method and device relates to the technical field of energy system load prediction, and comprises the steps of obtaining energy supply amount historical data, energy consumption historical data and influence factors related to the energy consumption historical data as input data; judging whether an abnormal value exists in the input data or not; if not, outputting the input data to the short-term load prediction model; if yes, removing the abnormal value, performing data filling and repairing on the removed abnormal value, and outputting the repaired input data to the short-term load prediction model; the short-term load prediction model selects data in a corresponding time period according to a time period needing to be predicted; performing normalization processing on the selected data; according to the normalized data, an extreme learning machine based on a simulated annealing algorithm is used to calculate and obtain a load prediction result; the method is used for solving the problem that a short-term load prediction result of an agricultural park integrated energy system is inaccurate.

Description

Technical Field

[0001] The present invention relates to the technical field of energy system load forecasting, and in particular to a method and device for short-term load forecasting of an integrated energy system for an agricultural park. Background Art

[0002] The context for a low-carbon, green energy transition is rooted in the harsh reality of global climate change and the urgent need for sustainable development. With the acceleration of industrialization, the intensive use of fossil fuels has led to rising atmospheric greenhouse gas concentrations, triggering a series of environmental problems, including global warming, extreme weather events, and ecosystem damage. To address these challenges, the international community has reached a series of agreements and goals aimed at limiting the rise in global average temperatures and promoting action by countries to reduce greenhouse gas emissions. Within this context, energy transition has become crucial, requiring a shift from reliance on high-carbon energy sources such as coal, oil, and natural gas to clean and sustainable energy sources such as wind, solar, hydro, and nuclear power.

[0003] Focusing on the agricultural sector, modern agricultural parks often feature high technological content, high conversion rates, and the integration of multiple factors to ensure the quality and efficiency of agricultural products. This, in turn, leads to high energy consumption and high emissions. Specifically, modern agricultural parks often integrate factors such as crop cultivation, livestock and poultry farming, aquaculture, agricultural product processing, and cold storage. Based on this, a large number of production environment temperature control equipment, large-scale agricultural production machinery, and monitoring and control devices at various levels are put into operation, all of which consume a large amount of energy and resources. To implement green, low-carbon, efficient, and sustainable energy systems in agricultural scenarios, achieve efficient temperature control of the production environment, low-carbon energy use for agricultural machinery, resource utilization of agricultural waste, and low-carbon energy use for agricultural product processing and storage, the introduction of green, low-carbon, integrated energy systems and the simultaneous circular economy of materials and energy in agricultural parks are important means to promote agricultural modernization, drive the development of related industries, promote scientific and technological innovation and industrial upgrading, and reduce negative environmental impacts.

[0004] An integrated energy system utilizes advanced physical information technology and innovative management models to integrate multiple energy resources within a specific region, including coal, oil, natural gas, electricity, and thermal energy, to form an integrated energy system. The core goal of this system is to effectively improve energy efficiency and promote sustainable energy development while meeting diverse energy needs.

[0005] The efficient and stable operation of each energy subsystem within the integrated energy system, the coordinated operation between energy subsystems, and the formulation and implementation of integrated system scheduling strategies all depend on the integrated energy system monitoring and management platform. One of the important criteria of the integrated energy monitoring platform is the energy load forecast result. The timeliness and accuracy of the energy load forecast result not only directly affect the economic and efficient allocation and utilization of energy in the integrated energy system, but also directly affect the reliability of the integrated energy system operation. Therefore, energy load forecasting is one of the key technologies to realize the intelligent, automated and sustainable development of the integrated energy system.

[0006] Existing load forecasting methods for integrated energy systems primarily rely on statistical analysis and machine learning. These methods have three drawbacks. First, their effectiveness is limited: Statistical analysis can only determine the relationship between independent and dependent variables based on statistical principles and establish a correlation regression equation, but this approach suffers from poor adaptability and nonlinear fitting capabilities. Machine learning methods, such as support vector regression and Bayesian networks, are unable to support large, potentially expanding datasets and, consequently, cannot meet the load forecasting needs of large, complex integrated energy systems. Second, the accuracy of multivariate load forecasting is low: Some researchers have gradually recognized the shortcomings of using statistical analysis and machine learning methods for load forecasting and have attempted to employ deep learning for load forecasting. However, these deep learning methods typically employ a simple overlay of feature sets and prediction models for a single load type to form a multivariate load forecast. This fails to meet the requirements for extracting correlation features under the synergistic effects of multiple loads, ultimately resulting in poor prediction results. Third, there are limitations in the applicable scenarios: although the research objects of existing load forecasting are named integrated energy systems, the system energy input categories are single and the load types are mostly combinations of electrical loads, cooling loads, and heating loads. The resource utilization of agricultural waste and the consumption demand for natural gas and biogas in the agricultural park scenario are not considered.

[0007] The Chinese patent with publication number CN118779664A discloses a method, device and equipment for multi-element load prediction of an integrated energy system. The method first obtains historical data of the multi-element loads of the integrated energy system and influencing factor data of the corresponding historical period, wherein the multi-element loads include electric load, cooling load and heating load, and the influencing factor data include at least one of meteorological condition data and holiday attribute data. Then, based on the correlation between the multi-element loads and between the multi-element loads and different influencing factor data, the specific data composition form of the sample data to be used is selected, which is also the data to be predicted when the multi-element load prediction is performed subsequently. Therefore, the multi-element load prediction model is used to perform multi-element load prediction of the integrated energy system while fully considering the correlation and coupling between the multi-element loads and between the multi-element loads and different influencing factor data, thereby improving the multi-element load prediction accuracy. However, this method lacks the integration of agricultural-specific factors, resulting in the model being unable to effectively capture the multi-element load characteristics of the agricultural integrated energy system, and the prediction accuracy is limited.

[0008] Therefore, we propose a prediction method for agricultural parks with accurate prediction results. Summary of the Invention

[0009] The purpose of the present invention is to provide a short-term load forecasting method and device for an integrated energy system in an agricultural park, which is used to solve the problem of inaccurate short-term load forecasting results of an integrated energy system in an agricultural park.

[0010] The present invention is achieved through the following technical solutions:

[0011] A short-term load forecasting method for an integrated energy system in an agricultural park, specifically comprising:

[0012] Based on a time series consisting of dates and times, historical energy supply data of each energy supply circuit in the agricultural park, historical energy consumption data of each energy load circuit, and influencing factors related to the historical energy consumption data are obtained as input data;

[0013] Use the integrated energy system load forecasting platform to determine whether there are outliers in the input data;

[0014] If not, the input data is output to the short-term load forecasting model;

[0015] If there are any, the outliers are removed, and the data of the removed outliers is filled and repaired, and then the repaired input data is output to the short-term load forecasting model;

[0016] The short-term load forecasting model selects data within the corresponding time period according to the time period to be predicted;

[0017] Normalize the selected data;

[0018] According to the normalized data, the load forecast results are calculated using the extreme learning machine based on the simulated annealing algorithm.

[0019] Furthermore, the outlier judgment process of the input data is as follows:

[0020] Using the data graphical processing system of the integrated energy system load forecasting platform, the input data are sorted in ascending chronological order, and the upper quartile F1, median F2, and lower quartile F3 of the sorted series are calculated;

[0021] The upper and lower margins were calculated based on the upper quartile F1, the median F2, and the lower quartile F3;

[0022] Data that is smaller than the lower edge or larger than the upper edge are treated as outliers and removed.

[0023] Furthermore, the steps of filling and repairing the data of the removed outliers and the data before and after them are as follows:

[0024] Determine whether the abnormal data is continuously missing;

[0025] If not, the arithmetic mean of the two moments before and after the abnormal value is used to fill the gap;

[0026] If so, select two adjacent dates with similar influencing factors to the outlier and calculate the arithmetic mean of the corresponding time data of the two adjacent dates for filling.

[0027] Furthermore, the short-term load forecasting model performs short-term load forecasting for the next w hours, selecting input data within w*24 hours.

[0028] Furthermore, the specific steps of normalizing the selected data are:

[0029] Extract m sample training data from the data set after outlier judgment, where each sample includes data for w*24 hours before the load forecast, and construct a training input matrix. Extract influencing factors related to the historical data of energy consumption for w hours after the load forecast, and construct a training output matrix.

[0030] From the data set after outlier judgment, a loop statement is used to extract n sample training data, where each sample includes data w*24 hours before the load forecast, to build a prediction input matrix, and extract the influencing factors related to the historical data of energy consumption w hours after the load forecast to build a label data matrix;

[0031] Call the minimum-maximum normalization library function to normalize the training input matrix, training output matrix, prediction input matrix, and label data matrix respectively, scale the matrix data to the range of [0,1], and obtain the normalized results of the training input, training output, prediction input, and label data.

[0032] Furthermore, the default values of the time variable w, the quantity variable m, and the quantity variable n are 1, 1200, and 400, respectively.

[0033] Furthermore, the calculation steps of the load forecast result are:

[0034] Define the number of influencing factors related to historical energy consumption data as the total number of variables x, and then generate an initial population X. Each individual in the population contains x variables, and the variables are randomly initialized in the range of [-2, 2].

[0035] Iteratively calculate the fitness of each individual in the initial population and output the optimal fitness;

[0036] Calculate the output layer weight matrix corresponding to the optimal fitness and the normalized load forecast results;

[0037] The normalized load forecast result is restored to obtain the load forecast result.

[0038] Furthermore, the fitness of each individual in the initial population is iteratively calculated, and the specific steps are as follows:

[0039] Before each iteration, an all-zero matrix DX is initialized to store the change of each individual in each iteration;

[0040] Calculate the update direction and distance of each individual in the population, recorded as the increment matrix;

[0041] For each individual i (i.e., row i, i∈(1,N)), calculate the fitness difference between it and other individuals j (i.e., column j, j∈(1,N)), and update the increment matrix;

[0042] Generate new candidate solutions based on the incremental matrix;

[0043] Set the input layer weight matrix, sig activation function, output layer weight matrix, hidden layer input value matrix and hidden layer output matrix, as well as the input layer bias matrix generated according to the new candidate solution to build a neural network model;

[0044] Using the neural network model, the output data corresponding to the predicted input matrix is calculated;

[0045] Compare the normalized results of the output data and label data corresponding to the prediction input matrix, and calculate the ratio of the number of samples with correct prediction results to the total number of samples as the performance indicator of the prediction algorithm;

[0046] Calculate the accuracy rate based on the prediction algorithm performance index, that is, the current best fitness;

[0047] In subsequent iterations, if the new candidate solution has a higher fitness, it will replace the current best fitness until the iteration ends.

[0048] Furthermore, the output layer weight matrix corresponding to the optimal fitness is used in combination with the historical data of energy consumption of each energy load circuit to calculate the multivariate load forecast data.

[0049] A short-term load forecasting device for an integrated energy system for an agricultural park includes a park energy supply circuit installation and collection terminal, a park energy load circuit installation and collection terminal, and a park energy system centralized control room. The park energy supply circuit installation and collection terminal is used to collect energy supplied by energy sources of each user in the integrated energy system, where energy sources include natural gas, self-produced biogas, solar energy, wind energy, electricity, and heat energy. The energy supply circuit includes an electric / cooling / heat output circuit of a gas generator set, an electric power output circuit of a photovoltaic generator set, an electric power output circuit of a wind turbine generator set, and a power supply circuit of a distribution network. The energy supply circuit collection terminal specifically includes a gas meter, an electric power meter, a heat meter, and a steam meter.

[0050] The energy load circuit of the park is installed with a collection terminal for collecting the energy consumed by the energy load of each user in the integrated energy network, where the load includes gas load, cooling and heating load, and electricity load; the load circuit includes the energy receiving circuit of the key production equipment of the park users, the energy receiving circuit of the high-energy-consuming equipment, the output circuit of the power distribution cabinet of the production workshop, the output circuit of the cylinder of the production workshop, the output circuit of the cooling and heating main line, and the output circuit of the power distribution box of the non-production building; the specific form of the load circuit collection terminal includes a gas meter, an electricity meter, a heat meter, and a steam meter;

[0051] The park energy system control room includes an integrated energy system load forecasting platform, which is used to collect, process and store data output by two acquisition terminals, calculate and display load forecast results.

[0052] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0053] The present invention discloses a method and device for short-term load forecasting of an integrated energy system for an agricultural park, which is conducive to improving input data by collecting multiple data, thereby improving the accuracy of subsequent forecast results;

[0054] Furthermore, optimizing extreme learning machines (ELMs) using a simulated annealing algorithm can improve load forecasting accuracy. A key feature of this algorithm is that it randomly initializes the connection weights from the input layer to the hidden layer, eliminating the need for adjustment. This random initialization simplifies model training because it eliminates the need for backpropagation to adjust the hidden layer weights, significantly reducing training complexity and computational complexity. Furthermore, once the weights from the input layer to the hidden layer are randomly initialized, the output layer weights can be directly calculated using simple mathematical operations (usually matrix operations). This single forward pass makes ELM training very fast when processing large datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flow chart of a method of the present invention;

[0056] Figure 2 This is a sample example diagram in an embodiment of the present invention;

[0057] Figure 3 This is a diagram showing the SABO-ELM prediction results for the next hour;

[0058] Figure 4 This is a diagram showing the SABO-ELM prediction results for the next 24 hours. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0060] As attached Figure 1 The short-term load forecasting method for an integrated energy system in an agricultural park is shown, specifically including:

[0061] Based on a time series consisting of dates and times, historical energy supply data of each energy supply circuit in the agricultural park, historical energy consumption data of each energy load circuit, and influencing factors related to the historical energy consumption data are obtained as input data;

[0062] Use the integrated energy system load forecasting platform to determine whether there are outliers in the input data;

[0063] If not, the input data is output to the short-term load forecasting model;

[0064] If there are any, the outliers are removed, and the data of the removed outliers is filled and repaired, and then the repaired input data is output to the short-term load forecasting model;

[0065] The short-term load forecasting model selects data within the corresponding time period according to the time period to be predicted;

[0066] Normalize the selected data;

[0067] Based on the normalized data, the load forecast results are calculated using the Extreme Learning Machine (ELM) based on the Simulated Annealing Based Optimization (SABO) algorithm.

[0068] As an embodiment, the process of determining abnormal values of input data is specifically as follows:

[0069] Using the data graphical processing system of the integrated energy system load forecasting platform, the input data are sorted in ascending chronological order, and the upper quartile F1, median F2, and lower quartile F3 of the sorted series are calculated;

[0070] The upper and lower margins were calculated based on the upper quartile F1, the median F2, and the lower quartile F3;

[0071] The upper edge calculation formula is:

[0072] F max =F1+1.5(F3-F1)

[0073] The formula for calculating the lower edge is:

[0074] F min =F3-1.5(F1-F3)

[0075] Data that is smaller than the lower edge or larger than the upper edge are treated as outliers and removed.

[0076] Furthermore, the steps of filling and repairing the data of the removed outliers and the data before and after them are as follows:

[0077] Determine whether the abnormal data is continuously missing;

[0078] If not, the arithmetic mean of the two moments before and after the abnormal value is used to fill the gap;

[0079] The calculation formula is:

[0080] f (t,d) =f (t-1,d)+ f (t+1,d) / 2

[0081] Where f (t,d) is the abnormal value at time t on day d, f (t-1,d) is the value at time t-1 on the dth day, f (t+1,d) is the value at time t+1 on the dth day;

[0082] If so, select two adjacent dates with similar influencing factors to the outlier and calculate the arithmetic mean of the corresponding time data of the two adjacent dates for filling;

[0083] The calculation formula is:

[0084] f (t,d) =f (t,d-n)+ f (t,d+n) / 2

[0085] Where f (t,d) is the abnormal value at time t on day d, f (t,d-n) is the value at time t on day dn, f (t,d+n) It is the value at time t on the d+nth day.

[0086] As an embodiment, the short-term load forecasting model performs short-term load forecasting for the next w hours, and selects input data within w*24 hours.

[0087] In addition, the specific steps of normalizing the selected data are:

[0088] Extract m sample training data from the data set after outlier judgment, where each sample includes w*24 hours of data before the load forecast, and construct the training input matrix input tra1n , and extract the influencing factors related to the historical data of energy consumption after w hours of load forecast, and construct the training output matrix output train ;

[0089] From the data set after outlier judgment, a loop statement is used to extract n sample training data, where each sample includes w*24 hours of data before the load forecast, and the prediction input matrix input is constructed. test , and extract the influencing factors related to the historical data of energy consumption after w hours of load forecast, and construct the label data matrix label test ;

[0090] Call the minimum-maximum normalization library function (mapminmax) to train the input matrix input train , training output matrix output train , prediction input matrix input test and label data matrix label testPerform normalization and scale the matrix data to the range of [0,1] to obtain the normalized result of the training input train _tn, normalized result of training output output train _tn, normalized result of prediction input test _tn and the normalized result label of the label data test _tn.

[0091] In particular, the default values of the time variable w, quantity variable m, and quantity variable n are 1, 1200, and 400, respectively. These values also support user-defined assignments; users can adjust these values to achieve day-ahead (24-72 hours) and intraday (4-12 hours) load forecasting.

[0092] As an embodiment, the steps for calculating the load forecast result are:

[0093] First, the simulated annealing program assigns the population size N to 30 and the number of iterations T to 100. The above assignments also support user-defined assignments.

[0094] Then, the number of influencing factors related to the historical data of energy consumption is defined as the total number of variables x, and then an initial population X is generated. Each individual in the population contains x variables, and the variables are randomly initialized in the range of [-2, 2].

[0095] Iteratively calculate the fitness of each individual in the initial population and output the optimal fitness;

[0096] In addition, before each iteration, an all-zero matrix DX is initialized to store the change of each individual in each iteration;

[0097] Calculate the updated direction and distance of each individual in the population, recorded as the incremental matrix DX(i,d), and this update is based on the difference between the current solution and other randomly selected solutions, as well as the difference in their fitness;

[0098] For each individual i (i.e., row i, i∈(1,N)), calculate the fitness difference between it and other individuals j (i.e., column j, j∈(1,N)), and update the incremental matrix DX(i,d)d∈(1,x);

[0099] DX(i,d)=(X(j,d)-I*X(i,d))*sign(fit(i)-fit(j))

[0100] Where X(i, d) is the current solution, X(j, d) is another randomly selected solution, I is a random integer between 1 and 3, which is used to adjust the size of the difference, and sign(fit(i)-fit(j)) is the sign function, which can determine the direction of the update based on the difference in fitness.

[0101] Generate new candidate solutions based on the incremental matrix

[0102]

[0103] Where X(i,:) represents the selection of all elements in the i-th row of the matrix X, rand(1,x) is a randomly generated 1×x random floating point vector, where the value of each element is between 0 and 1, and DX(i,:) represents the selection of all elements in the i-th row of the matrix DX. Must meet both as well as The new solution is controlled to be within the range of [-4, 4]. In the above steps, because all fitness differences between different individuals in the population and between different load influencing factors are taken into account, the situation of simultaneous cold and hot loads in the production scenario of agricultural parks is included. In traditional load forecasting methods, cold and hot loads are supplied independently without considering the impact of mutual assistance, resulting in the unenergy-saving phenomenon of heating and cooling during the actual operation of the load. By considering the complementarity of cold and hot loads, the comprehensive energy load energy consumption of the park can be further optimized. In addition, through the iterative optimization of load influencing factors using the simulated annealing algorithm, the importance of features can be measured for feature selection, and irrelevant features can be gradually eliminated to reduce the dimension of the feature space, thereby improving the performance of the prediction model.

[0104] Set the input layer weight matrix IW, where the input layer weight matrix IW is The matrix and the initialization assignment of each element are between [-1,1], the sig activation function f(x), the output layer weight matrix LW, the hidden layer input value matrix tempH and the hidden layer output matrix H, and the input layer bias matrix B generated according to the new candidate solution. The input layer bias matrix B is randomly generated Matrix, copy along the column direction of matrix B times, generate a new Matrix BiasMatrix is used to build a neural network model;

[0105] tempH=IW*input train _tn+BiasMatrix

[0106] H=1 / (1+exp(-tempH))

[0107] And the output layer weight matrix LW is the inverse of the result matrix of the training set H T Multiply the output value matrix tempH of the hidden layer by the transposed matrix, that is:

[0108] LW=(H T ) -1 tempH T

[0109] Using the neural network model, the output data corresponding to the predicted input matrix is calculated;

[0110] Using the obtained output layer weight matrix LW, combined with the input data of the test set, the output data output corresponding to the test set can be solved test _tn, which is the normalized load forecast result, is calculated as follows:

[0111] output test _tn=LW / (1+exp(-(IW*input test _tn+BiasMatrix))

[0112] Because the weights of the output layer are learned by randomly initializing the connection weights from the input layer to the hidden layer, the residual is minimized, which can significantly improve the training speed and has better generalization performance.

[0113] In the above steps, the weights of the output layer are learned by randomly initializing the connection weights from the input layer to the hidden layer, thereby minimizing the residual error. This can significantly improve the training speed and have better generalization performance.

[0114] Compare the output data corresponding to the predicted input matrix output test _tn and the normalized result label of the label data test _tn, and calculate the proportion of the number of samples with correct prediction results to the total number of samples, which is used as the performance indicator of the prediction algorithm correct predictions ;

[0115] correct predictions =sum(output test _tn==label test _tn)

[0116] Calculate the accuracy based on the prediction algorithm performance index predictions , that is, the current best fitness best sofar ;

[0117] accuracy predictions =correct predictions / length(label test _tn)

[0118] In subsequent iterations, if the new candidate solution has a higher fitness, it will replace the current best fitness until the iteration ends.

[0119] Calculate the output layer weight matrix LW corresponding to the optimal fitness and the normalized load forecast result output test _tn;

[0120] The normalized load forecast result is restored to obtain the load forecast result.

[0121] In addition, the output layer weight matrix LW corresponding to the optimal fitness is used, combined with the historical data input of energy consumption of each energy load circuit history , calculate and obtain multivariate load forecast data;

[0122] output prediction =LW / (1+exp(-(IW*input history +BiasMatrix)).

[0123] It should be noted that the multivariate load forecast results calculated in the above steps are displayed to users through the integrated energy system load forecasting platform, and users can choose between table or graphic formats;

[0124] The following example verifies the accuracy of the SABO-ELM model's load forecasting training results. Historical data on electricity load, natural gas load, heat load, humidity, temperature, wind speed, pressure, precipitation visibility, vapor pressure (e), and perceived temperature (AT) are used as input samples for the load forecasting algorithm. Figure 2 For sample examples;

[0125] The SABO-ELM algorithm is used to forecast the electricity load, natural gas load, and heat load for the next 1 hour and 24 hours. When using the two optimization methods, the evolutionary generation needs to be set. By default, it is set to 100 (the higher the evolutionary generation, the smaller the optimization deviation, so a larger evolutionary generation is selected). Figure 3 and Figure 4 The results show that SABO-ELM outperforms the traditional ELM algorithm in terms of accuracy in forecasting for the next hour and 24 hours. This verifies the SABO-ELM algorithm's ability to construct a short-term thermal and electrical load forecasting model for a park, achieving day-ahead (24-72 hours) and intraday (4-12 hours) forecasts for various energy sources such as electricity, heat, and gas.

[0126] A short-term load forecasting device for an integrated energy system for an agricultural park includes a park energy supply circuit installation and collection terminal, a park energy load circuit installation and collection terminal, and a park energy system centralized control room. The park energy supply circuit installation and collection terminal is used to collect energy supplied by energy sources of each user in the integrated energy system, where energy sources include natural gas, self-produced biogas, solar energy, wind energy, electricity, and heat energy. The energy supply circuit includes an electric / cold / heat output circuit of a gas generator set, an electric power output circuit of a photovoltaic generator set, an electric power output circuit of a wind turbine generator set, and a power supply circuit of a distribution network. The park energy supply circuit installation and collection terminal includes a gas meter, an electric power meter, a heat meter, and a steam meter.

[0127] The energy load circuit of the park is installed with a collection terminal for collecting the energy consumed by the energy load of each user in the integrated energy network, wherein the load includes gas load, cooling and heating load, and electricity load; the load circuit includes the energy receiving circuit of the key production equipment of the park users, the energy receiving circuit of the high-energy-consuming equipment, the output circuit of the power distribution cabinet of the production workshop, the output circuit of the cylinder of the production workshop, the output circuit of the cooling and heating main line, and the output circuit of the power distribution box of the non-production building; and the specific form of the energy load circuit collection terminal of the park includes a gas meter, an electricity meter, a heat meter, and a steam meter;

[0128] The park energy system control room includes an integrated energy system load forecasting platform, which is used to collect, process and store data output by the two acquisition terminals, calculate and display the load forecast results. The displayed data provides a data basis for users to subsequently carry out efficient integrated energy scheduling within the park, ensure the safe operation of the integrated energy system, and plan internal and external energy trading strategies.

[0129] In addition, the device can realize the day-ahead (24-72 hours) and intraday (4-12 hours) forecast of various energy demands such as electricity, heat and gas, and display the forecast results to users in real time;

[0130] This device can make short-term predictions on the load size of each load node by combining the historical energy supply data of each power node and the historical energy consumption data of the load node in the integrated energy network with the main factors affecting the electricity, heat and gas consumption of the park, including meteorological factors (temperature, humidity, weather type, wind speed and solar radiation intensity, etc.), historical load data (electricity load, heat load and gas load), and user production factors (production plan, agricultural waste recycling volume, self-produced biogas volume, etc.).

[0131] This device can consider whether there is an impact between different loads and the degree of mutual influence, thus covering the situation where cold and hot loads coexist in the production scenario of agricultural parks. In traditional load forecasting methods, cold and hot loads are supplied separately without considering the impact of mutual assistance, resulting in the unenergy-saving phenomenon of heating and cooling during the actual operation of the load. By considering the complementarity of cold and hot loads, the comprehensive energy load energy consumption of the park can be further optimized;

[0132] This device can iteratively optimize the load influencing factors through the simulated annealing algorithm, measure the importance of features to perform feature selection, and gradually eliminate irrelevant features to reduce the dimension of the feature space, thereby improving the performance of the prediction model;

[0133] This device can learn the weights of the output layer by randomly initializing the connection weights from the input layer to the hidden layer, thereby minimizing the residual error. This can significantly improve the training speed and has good generalization performance.

[0134] This device also has a wide range of applications: the short-term load forecasting device for the integrated energy system only needs to collect the historical energy supply data of each power node in the network, the historical energy consumption data of the load nodes, and the main factors affecting the electricity, heat and gas consumption of the park entered by the user, including meteorological factors (temperature, humidity, weather type, wind speed and solar radiation intensity, etc.), historical load data (electricity load, heat load and gas load), and user production factors (production plan, agricultural waste recycling volume, self-produced biogas volume, etc.), to deduce and solve the load forecast results of each integrated energy network. The basic input data is easy to obtain and the application scenario compatibility is strong.

[0135] Consider the impact between the loads of the integrated energy system: During the actual operation of the integrated energy system, there are often working conditions where cold and hot loads exist at the same time, that is, heating is required on one side and cooling is required on the other side. If the heat generated by refrigeration can be moved to the heat load, the overall energy consumption of the system can be reduced. Therefore, the coordination between the loads of the integrated energy system is an important factor in optimizing the system energy consumption and conducting load forecasting at the same time. The short-term load forecasting device of the integrated energy system takes into account all fitness differences between different individuals in the population and between different load influencing factors, and fully considers the impact between the loads of the integrated energy system.

[0136] Highly user-friendly: The integrated energy system short-term load forecasting device uses the built-in data repair program of the integrated energy system load forecasting platform to identify and eliminate outliers in the collected data, and at the same time fills and repairs the data before and after the outliers are eliminated, without placing excessive demands on the accuracy of the data entered by the user.

[0137] The algorithm is simple and reliable: The optimized simulated annealing extreme learning machine algorithm built into the integrated energy system short-term load forecasting device has the key feature of randomly initializing the connection weights from the input layer to the hidden layer, and there is no need to adjust these weights. This random initialization method simplifies the model training process because there is no need to adjust the weights of the hidden layer through the backpropagation algorithm, which greatly reduces the complexity and amount of training. Moreover, once the weights from the input layer to the hidden layer are randomly initialized, the weights of the output layer can be directly calculated through simple mathematical operations (usually matrix operations). This single forward pass feature makes the ELM training speed very fast when processing large-scale data sets.

[0138] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A short-term load forecasting method for an integrated energy system in an agricultural park, characterized in that: Specifically include: Based on a time series consisting of dates and times, historical energy supply data of each energy supply circuit in the agricultural park, historical energy consumption data of each energy load circuit, and influencing factors related to the historical energy consumption data are obtained as input data; Use the integrated energy system load forecasting platform to determine whether there are outliers in the input data; If not, the input data is output to the short-term load forecasting model; If there are any, the outliers are removed, and the data of the removed outliers is filled and repaired, and then the repaired input data is output to the short-term load forecasting model; The short-term load forecasting model selects data within the corresponding time period according to the time period to be predicted; Normalize the selected data; According to the normalized data, the load forecast results are calculated using the extreme learning machine based on the simulated annealing algorithm.

2. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 1, characterized in that: The process of judging the outliers of input data is as follows: Using the data graphical processing system of the integrated energy system load forecasting platform, the input data are sorted in ascending chronological order, and the upper quartile F1, median F2, and lower quartile F3 of the sorted series are calculated; The upper and lower margins were calculated based on the upper quartile F1, the median F2, and the lower quartile F3; Data that is smaller than the lower edge or larger than the upper edge are treated as outliers and removed.

3. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 2, characterized in that: The steps for filling and repairing the data of the removed outliers and the data before and after them are as follows: Determine whether the abnormal data is continuously missing; If not, the arithmetic mean of the two moments before and after the abnormal value is used to fill the gap; If so, select two adjacent dates with similar influencing factors to the outlier and calculate the arithmetic mean of the corresponding time data of the two adjacent dates for filling.

4. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 1, characterized in that: The short-term load forecasting model performs short-term load forecasting for the next w hours, selecting input data within w*24 hours.

5. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 4, characterized in that: The specific steps of normalizing the selected data are as follows: Extract m sample training data from the data set after outlier judgment, where each sample includes data for w*24 hours before the load forecast, and construct a training input matrix. Extract influencing factors related to the historical data of energy consumption for w hours after the load forecast, and construct a training output matrix. From the data set after outlier judgment, a loop statement is used to extract n sample training data, where each sample includes data w*24 hours before the load forecast, to build a prediction input matrix, and extract the influencing factors related to the historical data of energy consumption w hours after the load forecast to build a label data matrix; Call the minimum-maximum normalization library function to normalize the training input matrix, training output matrix, prediction input matrix, and label data matrix respectively, scale the matrix data to the range of [0,1], and obtain the normalized results of the training input, training output, prediction input, and label data.

6. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 5, characterized in that: The default values of the time variable w, the quantity variable m, and the quantity variable n are 1, 1200, and 400, respectively.

7. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 1, characterized in that: The calculation steps of the load forecast result are: Define the number of influencing factors related to historical energy consumption data as the total number of variables x, and then generate an initial population X. Each individual in the population contains x variables, and the variables are randomly initialized in the range of [-2, 2]. Iteratively calculate the fitness of each individual in the initial population and output the optimal fitness; Calculate the output layer weight matrix corresponding to the optimal fitness and the normalized load forecast results; The normalized load forecast result is restored to obtain the load forecast result.

8. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 7, characterized in that: The iterative calculation of the fitness of each individual in the initial population is carried out in the following specific steps: Before each iteration, an all-zero matrix DX is initialized to store the change of each individual in each iteration; Calculate the update direction and distance of each individual in the population, recorded as the increment matrix; For each individual i (i.e., row i, i∈(1,N)), calculate the fitness difference between it and other individuals j (i.e., column j, j∈(1,N)), and update the increment matrix; Generate new candidate solutions based on the incremental matrix; Set the input layer weight matrix, sig activation function, output layer weight matrix, hidden layer input value matrix and hidden layer output matrix, as well as the input layer bias matrix generated according to the new candidate solution to build a neural network model; Using the neural network model, the output data corresponding to the predicted input matrix is calculated; Compare the normalized results of the output data and label data corresponding to the prediction input matrix, and calculate the ratio of the number of samples with correct prediction results to the total number of samples as the performance indicator of the prediction algorithm; Calculate the accuracy rate based on the prediction algorithm performance index, that is, the current best fitness; In subsequent iterations, if the new candidate solution has a higher fitness, it will replace the current best fitness until the iteration ends.

9. The method for short-term load forecasting of an integrated energy system for an agricultural park according to claim 8, characterized in that: The multivariate load forecast data is calculated by using the output layer weight matrix corresponding to the optimal fitness and combining the historical data of energy consumption of each energy load circuit.

10. A short-term load forecasting device for an integrated energy system for an agricultural park, which implements the short-term load forecasting method for an integrated energy system for an agricultural park as described in any one of claims 1 to 9, characterized in that: It includes a park energy supply circuit installation and collection terminal, a park energy load circuit installation and collection terminal, and a park energy system control room. The park energy supply circuit installation and collection terminal is used to collect the energy supplied by the energy source of each user in the integrated energy system, where the energy includes natural gas, self-produced biogas, solar energy, wind energy, electricity, and heat energy; the energy supply circuit includes the gas generator set electricity / cooling / heat output circuit, the photovoltaic generator set power output circuit, the wind turbine generator set power output circuit, and the distribution network power supply circuit; The energy load circuit of the park is equipped with a collection terminal for collecting the energy consumed by the energy load of each user in the integrated energy network, where the load includes gas load, cooling and heating load, and electricity load; the load circuit includes the energy receiving circuit of the key production equipment of the park users, the energy receiving circuit of the high-energy-consuming equipment, the output circuit of the power distribution cabinet of the production workshop, the output circuit of the cylinder of the production workshop, the output circuit of the cooling and heating main line, and the output circuit of the power distribution box of the non-production building; The park energy system control room includes an integrated energy system load forecasting platform, which is used to collect, process and store data output by two acquisition terminals, calculate and display load forecast results.

Images