A comprehensive energy system multi-element load prediction method based on a load participation factor

Abstract

The application discloses a kind of based on load participation factor comprehensive energy system multivariate load prediction method, the method includes the following steps: data acquisition and pretreatment;Calculate load participation factor;Replace load data;Construct sample sequence;Build multivariate load prediction model;Multivariate load prediction.The present application extracts the hidden information between multivariate load and overall load through load participation factor, establishes neural network model to excavate the coupling characteristics between multivariate load, and improves the accuracy of multivariate load prediction.In addition, load participation factor application criteria are proposed, which improves the generalization ability of the prediction method in different application scenarios.

Description

Technical Field

[0001] This invention relates to the field of integrated energy system load forecasting, and in particular to an integrated energy load forecasting method based on load participation factors. Background Technology

[0002] Integrated energy systems, with the power system at their core, achieve "multi-source complementarity" of various energy sources with different physical characteristics, such as electricity, cooling, and heating, at a horizontal level, and a high degree of coordination between "source-grid-load-storage" at a vertical level. This facilitates the full utilization of renewable energy and the cascade utilization of non-renewable energy. Load forecasting, as the primary prerequisite for energy management and optimized scheduling in integrated energy systems, can economically and rationally arrange the conversion of various energy sources, maintain the safe and stable operation of the power grid, and improve economic and social benefits.

[0003] Currently, traditional load forecasting methods mainly consist of conventional methods, traditional machine learning methods, and cutting-edge machine learning methods. Conventional methods, primarily represented by time series analysis and regression analysis, while effective for short-term, single-load forecasting, struggle to establish accurate mathematical models and predict nonlinear data due to the increasing complexity of integrated energy systems and the combined influence of multiple factors on energy demand as integrated systems evolve. Traditional machine learning methods, such as decision trees, random forests, and support vector regression, achieve good forecasting results without requiring complex mathematical models. However, with the integration of renewable energy and demand-side response, integrated energy systems need to forecast loads with diverse physical characteristics simultaneously, pushing traditional machine learning methods to their limits. Cutting-edge machine learning methods, centered on deep learning, learn hidden features layer by layer through multi-layer nonlinear mapping. This not only provides powerful explanations for complex nonlinear structures but also achieves good forecasting results without requiring complex mathematical models.

[0004] However, current load forecasting models do not extract the hidden information between multiple loads and overall load demand, and most forecasts are only for individual energy loads, ignoring the coupling characteristics between various energy sources in the integrated energy system and severing the coupling relationship between energy systems. Summary of the Invention

[0005] To address the problems mentioned above, the technical solution used in this invention is: a multi-element load forecasting method for integrated energy systems based on load participation factors, comprising the following steps:

[0006] Step 1, Data Acquisition and Preprocessing: Obtain historical load data from multiple sources, unify the dimensions of the multiple load data and sum them up to obtain the total load data, and perform data preprocessing to obtain the preprocessed historical load data;

[0007] Step 2, Calculate the load participation factor: Calculate the load participation factor based on the preprocessed historical load data;

[0008] Step 3, Replace load data: Based on the load participation factor values ​​and the correlation between the multivariate loads and the total load, formulate load participation factor application criteria. Replace the load data that meets the criteria with the load participation factor to obtain the processed load data.

[0009] Step 4, construct the sample sequence: normalize the load data processed in Step 3, and select the time step using autocorrelation analysis to obtain the sample sequence.

[0010] Step 5, Construct a multivariate load forecasting model: Construct a multi-task learning neural network, input the sample sequence into the neural network for training, and obtain a multivariate load forecasting model;

[0011] Step 6, Multivariate load forecasting: The sample sequence is constructed using the same data processing methods as in steps 1 to 4 for the multivariate load data collected online. This sample sequence is then input into the multivariate load forecasting model, and the output data of the forecasting model is restored to obtain the multivariate load forecasting results.

[0012] The data preprocessing in step one above includes: filling in missing values ​​and removing outliers.

[0013] The specific method for calculating the load participation factor in step two above is as follows: for any load type at any time, the proportion of that type of load in the total load is the load participation factor.

[0014] The application criteria for the load participation factor in step three above are as follows:

[0015] (1) The absolute value of the correlation coefficient between this type of load and the total load is greater than ;

[0016] (2) The load participation factor value for this type of load is greater than .

[0017] in The set correlation threshold has a value range of [0.5, 0.95]. The set threshold for the load participation factor has a range of values. , The number of different types of loads.

[0018] If both of the above criteria are met simultaneously, then the load is considered to meet the load participation factor application criteria.

[0019] The specific steps for restoring the data in step six above are as follows:

[0020] (1) Perform inverse normalization on the multivariate load data and total load data output by the model;

[0021] (2) For load types that apply load participation factors, multiply the inverse normalized load data by the total load data to obtain the load forecast result; for load types that do not apply load participation factors, directly use the inverse normalized load data as the forecast result. Attached Figure Description

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0024] To further illustrate the features of the present invention, please refer to the following detailed description and accompanying drawings. The drawings are for reference and illustration only and are not intended to limit the scope of protection of the present invention.

[0025] like Figure 1 As shown, this invention employs a multi-factor load forecasting method for integrated energy systems based on load participation factors, comprising the following steps:

[0026] S1 Data Acquisition and Preprocessing: Acquire historical load data from multiple sources, unify the dimensions of the multiple load data and sum them up to obtain the total load data, and perform data preprocessing to obtain the preprocessed historical load data.

[0027] S2 calculates the load participation factor: Based on the preprocessed historical load data, the load participation factor is calculated;

[0028] S3 Replacement Load Data: Based on the load participation factor values ​​and the correlation between multivariate loads and total load, load participation factor application criteria are formulated. Load data that meets the criteria is replaced with load participation factors to obtain processed load data.

[0029] S4 Constructing the Sample Sequence: Normalize the load data processed in step three, and use autocorrelation analysis to select the time step to obtain the sample sequence.

[0030] S5 Constructing a multivariate load forecasting model: Construct a multi-task learning neural network, input the sample sequence into the neural network for training, and obtain a multivariate load forecasting model;

[0031] S6 Multivariate Load Forecasting: The same data processing methods as in steps one to four are used to construct sample sequences from the online collected multivariate load data. These sequences are then input into the multivariate load forecasting model, and the output data of the forecasting model is restored to obtain the multivariate load forecasting results.

[0032] Furthermore, S1 data acquisition and preprocessing includes the following sub-steps:

[0033] S11: Obtain historical load data from multiple sources;

[0034] S12: Unify historical load data from multiple sources to the same unit of measurement;

[0035] S13: For any given moment, sum the various load values ​​to obtain the total load value data for that moment;

[0036] S14: Fill in missing values ​​for the multivariate load history data;

[0037] S15: Perform outlier processing on the multivariate historical load data after missing value filling to obtain preprocessed historical load data.

[0038] This embodiment collects electricity, heat, and cooling load data from a university's integrated energy system from January 1, 2018 to December 31, 2019 as historical multi-load data in S11. The sampling frequency is 1 hour, and there are 17,520 data points for each type of load. The electricity, cooling, and heating load data at time t are E(t), C(t), and H(t), respectively.

[0039] Electricity, cooling, and heating loads have different units of measurement. In this dataset, electricity load is measured in kW, cooling load in Ton / hrs, and heating load in mmBtu / hr. To intuitively analyze various load data and calculate total load and load participation factor data, it is necessary to unify all loads to the same unit of measurement, which is kW in this case.

[0040] S12 uses the following formula to unify historical load data from multiple sources to the same unit of measurement:

[0041] 1kW=0.284Ton / hrs=0.0034mmBtu / hr

[0042] The electricity, cooling, and heating load data after unifying the units of measurement are as follows:

[0043] , , ,

[0044] Where T represents the historical record time.

[0045] S13 For any given time, sum the various load values ​​to obtain the total load value data for that time. The calculation formula is:

[0046] ,

[0047] The formula for filling missing values ​​in S14 multivariate load historical data is as follows:

[0048]

[0049] in The load data after filling , , These represent the load data for the hour before, the two hours before, and the day before, respectively, before the missing values ​​at time t.

[0050] After outlier handling by S15, preprocessed historical load data is obtained:

[0051] , , ,

[0052] Furthermore, the formula for calculating the load participation factor in S2 is as follows:

[0053]

[0054]

[0055]

[0056] in , , These correspond to the participation factors of electrical, cooling, and heating loads, respectively.

[0057] Furthermore, the application criteria for the load participation factor in S3 include the following two criteria:

[0058] Criterion 1: This type of load has a high correlation with the total load.

[0059]

[0060] in For total load and Pearson correlation coefficient between loads The high Pearson correlation threshold is set to 0.8 in this embodiment. The Pearson correlation coefficient is calculated as follows:

[0061]

[0062] The correlation coefficient ranges from -1 to 1, with a larger absolute value indicating a stronger correlation.

[0063] Criterion 2: This type of load accounts for a large proportion of the total load.

[0064]

[0065] in The number of load types is 3 in this dataset.

[0066] If both of the above criteria are met, then the load is considered to meet the load participation factor application criteria, and the load participation factor is used to replace multivariate load data for forecasting.

[0067] Table 1 shows the Pearson correlation coefficients between the three types of loads (electric, cooling, and heating) and the total load. When the absolute value of the correlation coefficient between the three types of loads and the total load is greater than 0.8, the load participation factor application criterion 1 is satisfied. Therefore, the electric and cooling loads satisfy the load participation factor application criterion 1, while the heating load does not.

[0068] Table 1. Pearson correlation coefficients between various types of loads and total load.

[0069]

[0070] Criterion 2 applies to the load participation factor. , far greater than Therefore, the electricity and cooling loads are met. As for the heating load... This does not meet the application criterion 2 for load participation factors.

[0071] Therefore, the electrical and cooling loads simultaneously meet both of the above-mentioned load participation factor application criteria, and the electrical and cooling load participation factors are used to replace the original electrical and cooling load data for prediction. However, the heating load does not meet either application criterion, and the heating load data is used directly for prediction. Therefore, the processed load data obtained in S3 is:

[0072] , , , ,

[0073] Furthermore, the S4 method for constructing sample sequences includes the following subdivision steps:

[0074] S31: Determine the sample sequence length based on autocorrelation analysis;

[0075] S32: Normalize the load sample and the total load sample to obtain the sample sequence.

[0076] Specifically, S31 determined the sample sequence length to be 15 based on autocorrelation analysis.

[0077] The formula for normalizing the load sample and total load sample in S32 is as follows:

[0078]

[0079] in The load data at time t is the normalized value. The maximum value in the sequence. This is the minimum value in the sequence. The normalized electrical, cooling, and heating loads, as well as the total load data, are as follows: , , , , .

[0080] Furthermore, S5 constructs a multi-task learning neural network, inputs the sample sequences into the multi-task learning neural network for training, and obtains a multivariate load prediction model, including the following sub-steps:

[0081] S51: Construct a multi-task learning neural network;

[0082] S52: Input the sample sequence before time t in the training samples into the neural network multi-task prediction model to obtain the prediction result at time t;

[0083] S53: Iteratively train the neural network multi-task prediction model until the prediction error reaches the set threshold, then stop training to obtain the load prediction model.

[0084] The specific structure for constructing a multi-task learning neural network in S51 is as follows:

[0085] First layer: Task input layer, where input tasks correspond to electrical load participation factor, cooling load participation factor, heating load, and total load, respectively. , , , .

[0086] The second layer is a fully connected layer, which connects to the four task input layers and has 32 neurons.

[0087] The third layer is an LSTM layer, which connects to four fully connected layers and has 96 neurons.

[0088] The fourth layer: the fusion layer, which integrates the four sub-tasks into one learning task.

[0089] The fifth layer is an LSTM layer connected to a fusion layer with 96 neurons.

[0090] The sixth layer is a fully connected layer that transforms the output of a single LSTM layer into four subtask outputs with a total of 96 neurons.

[0091] The seventh layer is a tiled layer that connects four fully connected layers, transforming multidimensional input into one dimension.

[0092] The eighth layer is the output layer, which outputs the electrical load participation factor, cooling load participation factor, heating load, and total load prediction value, respectively.

[0093] Specifically, the sample sequence before time t in S52 mainly includes time t. , , , , , , , Data on electricity and cooling load participation factors, as well as heat load and total load data.

[0094] Furthermore, the S6 multivariate load forecasting includes the following detailed steps:

[0095] S61: Construct a sample sequence from the online-collected multivariate load data using the same data processing methods as in steps one to three, and input it into the multivariate load prediction model;

[0096] S62: Perform inverse normalization on the model output to obtain the inverse normalized model output;

[0097] S63: Multiply the load participation factor sequence in the inverse normalized model output by the total load sequence to obtain the multivariate load prediction result.

[0098] Specifically, S61 will use the sample sequence , , , After inputting the data into the prediction model, the model output results are obtained: electricity and cooling load participation factors, heat load, and total load prediction. , , , .

[0099] Specifically, in S62, the model output is denormalized to obtain the denormalized electrical load participation factor, cooling load participation factor, heating load, and total load forecast. , , , .

[0100] The specific method for calculating electrical load and cooling load forecasts in S63 is as follows:

[0101]

[0102]

[0103] Therefore, the multi-element load forecasting results in S6 are as follows:

[0104] , ,

[0105] To further verify that the method provided in this invention has high accuracy in predicting multivariate load data, the MAPE of various load prediction results is calculated using the following formula:

[0106]

[0107] Calculations show that the MAPE values ​​for electricity, heat, and cooling loads predicted by the method provided in this invention are 2.2209, 3.0321, and 3.0953, respectively. The prediction model proposed in this invention is compared with commonly used machine learning algorithms (such as neural network models, random forests, progressive gradient regression trees, and support vector regression models) using the same prediction sample data. The results are shown in Table 3.

[0108] Table 2 Comparison of MAPE results for multi-element load forecasting

[0109]

[0110] The above results demonstrate that the method of the present invention effectively improves the accuracy of multivariate load forecasting.

Claims

1. A multi-task learning-based method for predicting the multi-element load of an integrated energy system, characterized in that, Includes the following steps: Step 1, Data Acquisition and Preprocessing: Obtain historical load data from multiple sources, unify the dimensions of the multiple load data and sum them up to obtain the total load data, and perform data preprocessing to obtain the preprocessed historical load data; Step 2, calculate the load participation factor: Based on the preprocessed historical load data, for any load type at any time, calculate the proportion of that type of load in the total load as the load participation factor; Step 3, Replace load data: Based on the load participation factor values ​​and the correlation between the multivariate loads and the total load, formulate load participation factor application criteria. Replace the load data that meets the criteria with the load participation factor to obtain the processed load data. Step 4, construct the sample sequence: normalize the load data processed in Step 3, and select the time step using autocorrelation analysis to obtain the sample sequence; Step 5, Construct a multivariate load forecasting model: Construct a multi-task learning neural network, input the sample sequence into the neural network for training, and obtain a multivariate load forecasting model; Step 6, Multivariate load forecasting: The sample sequence is constructed using the same data processing methods as in steps 1 to 4 for the multivariate load data collected online. This sample sequence is then input into the multivariate load forecasting model, and the output data of the forecasting model is restored to obtain the multivariate load forecasting results.

2. The multi-source load forecasting method for integrated energy systems according to claim 1, characterized in that, The data preprocessing in step one includes: filling in missing values ​​and removing outliers.

3. The integrated energy system multi-source load forecasting method according to claim 1, characterized in that, The application criteria for the load participation factor in step three are as follows: (1) The absolute value of the correlation coefficient between this type of load and the total load is greater than φ; (2) The load participation factor for this type of load is greater than δ; Where φ is the set correlation threshold and δ is the set load participation factor threshold; If both of the above criteria are met simultaneously, then the load is considered to meet the load participation factor application criteria.

4. The integrated energy system multi-source load forecasting method according to claim 1, characterized in that, Step six involves restoring the data as follows: (1) Perform inverse normalization on the multivariate load data and total load data output by the model; (2) For the load type of the load participation factor, multiply the inverse normalized load data by the total load data to obtain the load forecast result; For load types that do not apply load participation factors, the inversely normalized load data is used directly as the prediction result.

Images