Industrial park net load prediction method and system based on multi-layer perceptron

Abstract

The present application discloses a method and system for predicting the net load of an industrial park based on a multi-layer perceptron, comprising the following steps: dividing the historical industrial park operation data according to the audio collection area to obtain the regional historical operation data corresponding to the audio collection area; clustering the historical operation data of each area based on the audio fluctuation, and obtaining the audio-aging correlation with the clustering result; using the historical industrial park operation data, the historical distributed energy operation data and the audio-aging correlation training to obtain the industrial park net load prediction model; using the current industrial park operation data, the current distributed energy operation data, the industrial park load net prediction model and the audio-aging correlation to obtain the industrial park net load prediction data. The beneficial effects of the present application: there is no need to construct an aging model for each industrial equipment, simplifying the overall model structure, while improving the prediction efficiency and prediction accuracy of the industrial park load.

Description

Technical Field

[0001] The present application relates to the technical field of industrial park load prediction, and in particular to an industrial park net load prediction method and system based on a multi-layer perceptron. Background Art

[0002] In traditional power grid management, power distribution and load forecasting mainly rely on centralized power generation models and deterministic forecasting techniques. These models transmit electricity from a few large power plants to various power consumption points through high-voltage transmission networks, and the forecasts of power generation and load are relatively stable. This forecasting method is mainly based on the analysis of historical load data and assumes that power generation and demand have small fluctuations, so it can more accurately predict future load demand. However, with the gradual commissioning of distributed energy (such as solar photovoltaic and wind power generation), this centralized power generation and deterministic forecasting model faces challenges, especially in industrial parks, where traditional forecasting models cannot handle the instability of distributed energy and the volatility of demand.

[0003] However, the centralized prediction method is insufficient in complex industrial parks. The interference of industrial equipment aging on load prediction results is ignored in the prediction process. When there are a large number of industrial equipment in the industrial park, the superposition of load fluctuations caused by industrial equipment aging can easily cause large prediction errors. In the related technology, an aging model is constructed for each industrial equipment separately. Not only is the data volume too large, but the model is too complex. When predicting, too much aging data is input into the prediction model, the calculation efficiency is low, and the large amount of data transmission increases the risk of data loss.

[0004] Net load refers to the total power load in a specific area minus the power generation output of distributed energy (such as solar power stations, wind power plants, biomass power generation, etc.) in the area. If there is a deviation in the load data forecast for the industrial park, the net load forecast value will also be biased, resulting in deviations in the optimization of power resource allocation strategies, which will ultimately affect the actual power supply situation.

[0005] The patent "Load Forecasting Model, Method and System Based on Multimodal Time Series Learning", publication number: CN117937475A, publication date: April 26, 2024, specifically discloses a method of introducing multimodal time series learning, extracting key features from multi-source time series data, and establishing a load forecasting model that takes into account energy storage. The net load forecasting of the industrial park in this solution is less targeted, and when the equipment in the industrial park ages, the load forecasting error is large.

[0006] The patent "A short-term power load forecasting method, system, equipment and medium", publication number: CN118352991A, publication date: July 16, 2024, specifically discloses obtaining the input original power load power sequence, preprocessing the input original power load power sequence, and obtaining historical load data; performing data decomposition on the historical load data to obtain several subsequence components with different frequencies; constructing a feature extraction model, extracting features from the subsequence components according to the feature extraction model, and obtaining the main features of the time series data; constructing a load forecasting model, training the load forecasting model according to the main features, predicting the subsequence components according to the load forecasting model to obtain predicted values, and superimposing and reconstructing the predicted values ​​corresponding to several subsequences to obtain the prediction results. Similarly, the net load forecasting of the industrial park in this scheme is less targeted, and when the equipment in the industrial park ages, the load forecasting error is large. Summary of the invention

[0007] In order to solve the problem in the prior art that the net load prediction of industrial parks includes too much data and low computational efficiency due to equipment aging factors, the present application provides a method and system for predicting the net load of industrial parks based on a multi-layer perceptron. The method obtains audio data in a region through the difference between the operating sound of aged industrial equipment and the operating sound of non-aged industrial equipment, and uses this to construct an audio-aging correlation. The aging fluctuations of industrial equipment in the entire region are displayed through the audio fluctuations in the entire region, thereby eliminating the need to construct an aging model for each industrial equipment and simplifying the overall model structure. At the same time, the prediction efficiency and accuracy of the load in the industrial park are improved, providing reliable input for subsequent net load prediction of the industrial park, ensuring the accuracy of the net load prediction of the industrial park, and ensuring the accuracy and efficiency of power dispatching.

[0008] To achieve the above technical objectives, the present application provides a technical solution, which is a method for predicting the net load of an industrial park based on a multi-layer perceptron, comprising the following steps: S1: obtaining historical industrial park operation data, dividing the historical industrial park operation data according to the audio collection area, and obtaining the regional historical operation data corresponding to the audio collection area; S2: clustering the historical operation data of each region based on the audio fluctuation, and obtaining the audio-aging correlation with the clustering results; S3: using the historical industrial park operation data, the historical distributed energy operation data and the audio-aging correlation to train a neural network model based on a multi-layer perceptron to obtain an industrial park net load prediction model; S4: obtaining the current industrial park operation data and the current distributed energy operation data, and obtaining the industrial park net load prediction data with the current industrial park operation data, the current distributed energy operation data, the industrial park load net prediction model and the audio-aging correlation.

[0009] Furthermore, S1 also includes: acquiring historical industrial park operation data, performing analog-to-digital conversion on the historical industrial park audio data, and performing human voice screening processing on the historical industrial park audio data after the analog-to-digital conversion; dividing the historical industrial park operation data after human voice screening processing according to the audio collection area, and obtaining regional historical operation data corresponding to the audio collection area.

[0010] Furthermore, the human voice screening processing of the historical industrial park audio data after analog-to-digital conversion also includes: using spectrum analysis to analyze the frequency distribution of the historical industrial park audio data after analog-to-digital conversion, and performing a human voice screening according to the frequency band difference between the operating sound of industrial equipment and the human voice; and performing a secondary human voice screening based on the repetition number threshold and the repetition period threshold based on the historical industrial park audio data after the human voice screening.

[0011] Furthermore, S2 also includes: extracting features from historical operating data of each region to obtain audio frequency fluctuation features; clustering the audio frequency fluctuation features using a clustering algorithm to obtain clustering results corresponding to the audio frequency fluctuation features; for each clustering result, calculating the continuous operating time of the industrial equipment in the clustering result, and constructing an audio-aging correlation based on the time series relationship of the audio frequency fluctuation features and the continuous operating time.

[0012] Furthermore, the construction of the audio-aging correlation according to the temporal relationship of the audio fluctuation characteristics and the continuous operation time includes: obtaining the smooth weight of the audio fluctuation characteristics according to the corresponding relationship between the audio fluctuation characteristics and the continuous operation time; obtaining the sudden change weight of the audio fluctuation characteristics according to the continuous operation time of the industrial equipment in each clustering result; and obtaining the audio-aging correlation based on the smooth weight of the audio fluctuation characteristics and the sudden change weight of the audio fluctuation characteristics.

[0013] Furthermore, S3 includes: constructing fluctuation parameters according to the smooth weights of the audio fluctuation characteristics and the sudden change weights of the audio fluctuation characteristics based on the historical industrial park audio data and historical time series; using the historical industrial park operation data, historical distributed energy operation data and fluctuation parameters to train a neural network model based on a multi-layer perceptron to obtain an industrial park net load prediction model.

[0014] Furthermore, S3 also includes: constructing a probability prediction model based on a multilayer perceptron network architecture; constructing a quantile regression model based on quantile regression; constructing a neural network model based on a multilayer perceptron according to the probability prediction model and the quantile regression model; obtaining future covariates using audio-aging correlation, and training a neural network model based on a multilayer perceptron according to future covariates, historical industrial park operation data, and historical distributed energy operation data to obtain an industrial park net load prediction model.

[0015] Furthermore, the construction of the probability prediction model based on the multi-layer perceptron network architecture includes: constructing the probability prediction model based on the correlation between distributed energy output and load, and the correlation between environmental factors and load.

[0016] Furthermore, the construction of a quantile regression model based on quantile regression includes: constructing an input layer based on rotation position embedding; constructing a historical feature encoder and a future covariate encoder based on historical features and future covariates; constructing a feature fusion layer, a quantile suggestion network and a quantile value network; constructing a quantile regression model according to the input layer, the historical feature encoder, the future covariate encoder, the feature fusion layer, the quantile suggestion network and the quantile value network.

[0017] Furthermore, the construction of a neural network model based on a multilayer perceptron according to the probability prediction model and the quantile regression model also includes: constructing a quantile loss function and constraints based on the Huber function and the 1-Wasserstein metric; and constructing a neural network model based on a multilayer perceptron based on the probability prediction model, the quantile regression model quantile loss function and constraints.

[0018] Furthermore, S4 also includes: obtaining the current industrial park operation data, and filtering out the human voice from the current industrial park audio data in the current industrial park operation data; obtaining the fluctuation parameters within a preset time period using the current industrial park audio data after filtering out the human voice and the audio-aging correlation; outputting the distributed energy output prediction data within a preset time period based on the current distributed energy operation data and the distributed energy output prediction model; using the distributed energy output prediction data, the environmental data within the preset time period, and the fluctuation parameters within the preset time period as future covariates of the industrial park load net prediction model, and outputting the industrial park net load prediction data.

[0019] Furthermore, S4 also includes: constructing a periodic fluctuation curve based on historical repetitive cycles, and constructing a quantile constraint interval based on the periodic fluctuation curve; using distributed energy output forecast data, environmental data within a preset time period, and fluctuation parameters within a preset time period as future covariates of the industrial park load net forecast model, selecting the quantile of the industrial park load net forecast model with the quantile constraint interval, and outputting the industrial park net load forecast data.

[0020] Furthermore, S4 also includes: utilizing an automatic regression reasoning mechanism to perform rolling prediction of the industrial park net load data according to a preset time period based on the current industrial park operation data, the current distributed energy operation data, the industrial park load net prediction model, and the audio-aging correlation to obtain the industrial park net load prediction data.

[0021] Another technical solution provided by the present application is an industrial park net load prediction system based on a multi-layer perceptron, which is used to implement the method as described above, including: an audio processing module, which constructs an audio-aging correlation based on historical industrial park operation data; a model construction module, which constructs a neural network model based on a multi-layer perceptron, and obtains an industrial park net load prediction model based on the audio-aging correlation and the neural network model training; a fusion analysis module, which outputs the industrial park net load prediction data based on the current industrial park operation data and the current distributed energy operation data.

[0022] The beneficial effects of the present application are as follows: 1. According to the operating sound generated by the industrial equipment over time and the aging condition of the industrial equipment, the correlation between the operating sound and the aging condition is obtained, and then according to the correlation between the operating sound frequency and aging in the current area, when training the net load prediction model of the industrial park, by incorporating the aging condition, the training of the net load prediction model of the industrial park is made more in line with the actual operating conditions of the industrial equipment. When the current industrial park operating data is obtained, the aging degree of the current industrial equipment is estimated according to the current industrial park audio data, and then the output industrial park load prediction data is in line with the actual operating conditions of the current industrial equipment, which improves the accuracy of the net load prediction of the industrial park without the need to construct an aging model for each industrial equipment separately, simplifies the model, reduces the computing pressure, and replaces the integration of the industrial equipment aging prediction results with the regional audio information changes, reduces the data processing volume, and is more efficient, thereby ensuring the timeliness of the output of the net load prediction data of the industrial park.

[0023] 2. Perform secondary human voice screening based on the sound repetition count threshold and repetition period threshold in the audio data of the industrial park. Since the human voice in the industrial park is usually not repetitive and periodic, while the sound of industrial equipment operation is repetitive and periodic due to the periodicity of the industrial equipment operation, the human voice fragments that are not repetitive and periodic can be effectively screened out based on the repetition count and repetition period, thereby further improving the human voice screening effect. In addition, since the human voice screening is first performed based on the frequency band difference, the calculation amount of the secondary screening repetition count and repetition period is reduced, thereby improving the screening efficiency.

[0024] 3. The future covariates are obtained by using the audio frequency-aging correlation. The net load prediction model of the industrial park is obtained by training the neural network model based on the multilayer perceptron according to the future covariates, the historical industrial park operation data and the historical distributed energy operation data.

[0025] At this time, the audio-aging correlation reflects the intrinsic relationship between the change of audio characteristics of industrial equipment and equipment performance during the aging process. The audio-aging correlation is used to obtain future covariates of the quantile regression model to improve the accuracy of quantile selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1This is a flow chart of the industrial park net load prediction method based on a multi-layer perceptron in this application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application is further described in detail below in conjunction with the drawings and examples. It should be understood that the specific implementation method described here is only an optimal embodiment of the present application, which is only used to explain the present application and does not limit the scope of protection of the present application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0028] like Figure 1 As shown, as the first embodiment of the present application, the industrial park net load prediction method based on the multi-layer perceptron includes the following steps:

[0029] S1: Acquire historical industrial park operation data, divide the historical industrial park operation data according to the audio collection area, and obtain regional historical operation data corresponding to the audio collection area;

[0030] S2: Perform cluster analysis on the historical operation data of each region based on the sound frequency fluctuation, and use the clustering results to obtain the sound frequency-aging correlation;

[0031] S3: Using historical industrial park operation data, historical distributed energy operation data and sound frequency-aging correlation to train a neural network model based on a multi-layer perceptron to obtain an industrial park net load prediction model;

[0032] S4: Obtain the current industrial park operation data and the current distributed energy operation data, and obtain the industrial park net load prediction data using the current industrial park operation data, the current distributed energy operation data, the industrial park net load prediction model, and the sound frequency-aging correlation.

[0033] Since there are a large number of industrial equipment in the industrial park, and the industrial equipment will generate operating sounds during operation, the correlation between the operating sounds and the aging conditions is obtained according to the operating sounds generated by the industrial equipment over time and the aging conditions of the industrial equipment. Then, according to the correlation between the operating sound frequency and aging in the current area, when training the net load prediction model of the industrial park, the aging conditions are included to make the training of the net load prediction model of the industrial park more in line with the actual operation conditions of the industrial equipment. When the current industrial park operation data is obtained, the aging degree of the current industrial equipment is estimated according to the current industrial park audio data. Then, the output industrial park load prediction data is in line with the actual operation conditions of the current industrial equipment, which improves the accuracy of the net load prediction of the industrial park without the need to build an aging model for each industrial equipment separately. The model is simplified, the calculation pressure is reduced, and the regional audio information changes are used instead of the industrial equipment aging prediction results integration. The data processing volume is reduced and the efficiency is higher, ensuring the timeliness of the output of the net load prediction data of the industrial park.

[0034] The industrial park operation data at least includes the industrial park load data, the industrial park audio data and the industrial park meteorological data. The regional historical operation data at least includes the regional audio data, the regional load data and the regional meteorological data. The audio collection area can be set according to the actual audio collection device setting area, or it can be set according to the type of factory in the industrial park. For example, in an industrial park, the type of industrial equipment installed in a factory building is generally fixed, so there is an audio collection point corresponding to each factory building, or there is an audio collection point corresponding to a production party to ensure the accuracy of the audio data benchmarking.

[0035] In addition, it should be noted that the industrial park operation data involved in this application (including but not limited to industrial park equipment information, industrial park load data, industrial park audio data, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0036] At this time, step S1 includes:

[0037] Obtain historical industrial park operation data and pre-process the historical industrial park operation data;

[0038] The pre-processed historical industrial park operation data is divided according to the audio collection area to obtain regional historical operation data corresponding to the audio collection area.

[0039] Among them, the preprocessing of historical industrial park operation data at least includes:

[0040] The audio data of the historical industrial park is converted into digital form, and the audio data of the historical industrial park after the analog-to-digital conversion is processed by filtering out human voices.

[0041] Although the operating noise of industrial equipment in industrial parks is loud, human voice data will interfere with the recognition of industrial equipment sound. Therefore, after the analog-to-digital conversion of the historical industrial park audio data, human voice screening is performed, and then the historical industrial park operation data after human voice screening is divided according to the audio collection area to ensure the accuracy of industrial equipment operation status recognition.

[0042] Specifically, the human voice screening process of the historical industrial park audio data after analog-to-digital conversion includes:

[0043] Spectrum analysis is used to analyze the frequency distribution of historical industrial park audio data after analog-to-digital conversion, and human voices are screened out based on the frequency band differences between the operating sounds of industrial equipment and human voices.

[0044] Since there are frequency band differences between human voice and the operating sound of industrial equipment, for example, the sound pressure level of human voice is lower while the sound pressure level of the operating sound of industrial equipment is higher, human voice can be screened out based on the frequency band differences between the operating sound of industrial equipment and human voice.

[0045] In other cases, the human voice screening process of the analog-to-digital converted historical industrial park audio data also includes:

[0046] Analyze the frequency distribution of the historical industrial park audio data after analog-to-digital conversion using spectrum analysis, and filter out human voices based on the frequency band differences between the sounds of industrial equipment operation and human voices;

[0047] According to the historical industrial park audio data after the first human voice screening, a second human voice screening is performed based on the repetition number threshold and the repetition cycle threshold.

[0048] The audio data of the historical industrial park is converted into digital signals using tools such as analog-to-digital converters for subsequent processing. Spectral analysis tools (such as fast Fourier transform, etc.) are then used to convert the audio signal from the time domain to the frequency domain. The main frequency bands of human voices and industrial equipment operation sounds are identified based on the spectrum diagram, and then a filter based on the frequency band differences of industrial equipment operation sounds and human voices is constructed. The filter is used to filter out human voices from the audio data of the historical industrial park. Human voices with obvious frequency band differences from the operation sounds of industrial park equipment can be filtered out by filtering out human voices through frequency band differences. However, since human voices and industrial park operation sounds may also have sounds in the same frequency band, secondary human voice screening is performed based on the sound repetition number threshold and repetition period threshold in the industrial park audio data. Since human voices in industrial parks are usually not repeatable and periodic, and the sound of industrial equipment operation is affected by the periodicity of industrial equipment operation and has repetitiveness and periodicity, human voice segments that are not repeatable and periodic can be effectively screened out based on the number of repetitions and repetition periods, thereby further improving the human voice screening effect. In addition, since human voice screening is first performed based on the frequency band difference, the number of repetitions and the calculation amount of the repetition period of secondary screening are reduced, and the screening efficiency is improved. It can be understood that although the operating sound of industrial equipment will change when it ages, the operating sound during aging also has repetitiveness and periodicity within a certain period of time and will not be mistakenly deleted. In this embodiment, the repetition number threshold and the repetition period threshold can be obtained based on expert experience, such as setting the repetition number threshold to three times and the repetition period threshold to 2. Only when the audio is repeated three times and there are two repetition periods, the audio is considered to belong to the industrial equipment audio. In some embodiments, the repetition period threshold may also be set to the industrial equipment maintenance period. When the audio frequency is periodically repeated, the audio frequency is considered to belong to the industrial equipment audio frequency only when the repetition period meets the industrial equipment maintenance period.

[0049] In some other cases, step S1 further includes:

[0050] Acquire historical industrial park operation data, and annotate historical industrial park audio data in the historical industrial park operation data with human voices and industrial equipment operation sounds;

[0051] Construct an initial model for voice screening based on spectrum analysis and repetition, and use the historical industrial park audio data after sound annotation to train the voice screening model to obtain a voice screening model;

[0052] The historical industrial park operation data after filtering out human voices is divided according to the audio collection area to obtain the regional historical operation data corresponding to the audio collection area.

[0053] Correspondingly, step S4 also includes:

[0054] Obtain the current industrial park operation data and the current distributed energy operation data, use the human voice screening model to screen out the human voice data in the current industrial park operation data, and obtain the current industrial park audio data;

[0055] The net load prediction data of the industrial park is obtained by using the current industrial park operation data including the current industrial park audio data, the current distributed energy operation data, the industrial park net load prediction model and the audio-aging correlation.

[0056] In this case, a human voice screening model corresponding to the first human voice screening and the second human voice screening is constructed, and the human voice screening model is trained through the historical industrial park audio data after sound annotation, and the repetition number threshold, repetition period threshold and screening frequency band threshold are obtained through training, so as to screen out the audio data less than the repetition number threshold, repetition period threshold and screening frequency band threshold, so that when the current industrial park operation data is obtained, the human voice data in the current industrial park operation data can be adaptively screened, and the audio data in different tenses and different regions can be targeted for human voice screening, so as to improve the accuracy of audio data processing. It is understandable that environmental noise other than human voice is also sporadic, and can also be screened out according to the number of repetitions and repetition periods.

[0057] The historical industrial park operation data of each audio collection area are screened for human voices according to the targeted number of repetitions, repetition cycles and screening frequency bands to obtain the regional historical operation data corresponding to the audio collection area. The historical operation data of each area are then analyzed to obtain the audio-aging correlation of each area.

[0058] In this embodiment, step S2 further includes:

[0059] Extract features from historical operation data of each region to obtain audio frequency fluctuation features;

[0060] Clustering is performed according to the audio frequency fluctuation characteristics by using a clustering algorithm to obtain clustering results corresponding to the audio frequency fluctuation characteristics;

[0061] For each clustering result, the continuous operation time of the industrial equipment in the clustering result is calculated, and the audio frequency-aging correlation is constructed according to the time series relationship of the audio frequency fluctuation characteristics and the continuous operation time.

[0062] Extract the characteristics with fluctuation from the regional historical audio data in the historical operation data of each region, such as frequency domain characteristics, time domain characteristics and audio change amount characteristics. Perform fast Fourier transform (FFT) on the audio signal to obtain spectrum information, calculate the frequency domain characteristics, calculate the maximum value, minimum value, mean value, standard deviation of the sound pressure level, peak factor (ratio of peak value to effective value), kurtosis (describing the degree of sharpness of signal distribution), etc. as time domain characteristics, calculate the rate of change of the audio signal in unit time as the audio change amount characteristics, cluster according to the audio fluctuation characteristics, and take the regional historical operation data with the same audio fluctuation characteristics as a cluster cluster, that is, the clustering result. At this time, for each clustering result, calculate the continuous operation time of the industrial equipment in the clustering result, and obtain the aging weight of the industrial equipment corresponding to each audio fluctuation characteristic. If the continuous operation time of the industrial equipment corresponding to an audio fluctuation characteristic is short, it means that the industrial equipment will mutate to the next aging degree faster under the audio fluctuation characteristic. From this, the audio fluctuation characteristics under different aging degrees and the continuous operation time of the industrial equipment under the aging state are obtained by the temporal relationship of the audio fluctuation characteristics and the continuous operation time.

[0063] Since the analog-to-digital conversion and human voice screening processing have been performed on the historical industrial park audio data in all historical industrial park operation data in S1, the spectral analysis of the audio data has been performed during the human voice screening process, and the analyzed internal features can be directly pre-stored, and the pre-stored feature information can be directly called when performing feature extraction.

[0064] Among them, the audio frequency-aging correlation is constructed according to the temporal relationship of the audio frequency fluctuation characteristics and the continuous operation time, including:

[0065] According to the corresponding relationship between the audio frequency fluctuation characteristics and the continuous running time, the audio frequency fluctuation characteristic smoothing weight is obtained;

[0066] The mutation weight of the acoustic frequency fluctuation feature is obtained according to the continuous operation time of the industrial equipment in each clustering result;

[0067] The audio-aging correlation is obtained based on the smooth weight of the audio frequency fluctuation feature and the sudden weight of the audio frequency fluctuation feature.

[0068] The smooth weight of the audio frequency fluctuation feature indicates the aging weight of the industrial equipment maintaining a stable audio frequency, and the sudden weight of the audio frequency fluctuation feature indicates the aging weight of the industrial equipment when the audio frequency suddenly changes. At this time, step S3 includes:

[0069] According to the historical industrial park audio data and historical time series, the fluctuation parameters are constructed according to the smooth weight of the audio fluctuation characteristics and the sudden weight of the audio fluctuation characteristics;

[0070] The net load prediction model of the industrial park is obtained by training the neural network model based on multi-layer perceptron using the historical industrial park operation data and fluctuation parameters.

[0071] If the load data of this area in time series A is B, the weight of the smoothing characteristic of the audio frequency fluctuation is , the mutation weight of the audio frequency fluctuation feature is ,get:

[0072] ;

[0073] ;

[0074] Among them, b is the fluctuation parameter, t1 is the time from the last change of the audio fluctuation feature of the A time series, m is the aging impact benchmark value, and y is the load prediction value under normal operation. Each change of the audio fluctuation feature is used as a segmentation. For example, the audio fluctuation feature mutation weight of the third audio fluctuation feature change is greater than the audio fluctuation feature mutation weight of the first audio fluctuation feature change. If the industrial equipment is maintained or replaced after the third audio fluctuation feature change, the audio fluctuation feature mutation weight of the fourth audio fluctuation feature is less than the audio fluctuation feature mutation weight of the third audio fluctuation feature change. In this way, the maintenance and replacement of industrial equipment is included in the load prediction consideration, further improving the accuracy of load prediction. Even when the audio fluctuation feature does not change, the industrial equipment is constantly aging, so the aging state of the industrial equipment in the stable period is included with the audio fluctuation feature gentle weight. During the training process, the aging impact benchmark value of the industrial equipment training corresponding to each area is obtained to improve the accuracy of the final net load prediction result.

[0075] Furthermore, the current industrial park operation data and the current distributed energy operation data are obtained, and the industrial park load forecast data is obtained by using the current industrial park operation data, the industrial park net load forecast model and the sound frequency-aging correlation. The industrial park net load forecast data is obtained based on the difference between the current distributed energy operation data and the industrial park load forecast data, providing a reference for the scheduling of traditional power generation resources.

[0076] As a second embodiment of the present application, step S3 further includes:

[0077] Construct a probabilistic prediction model based on a multi-layer perceptron network architecture;

[0078] Construct a quantile regression model based on quantile regression;

[0079] Construct a neural network model based on multilayer perceptron according to the probability prediction model and quantile regression model;

[0080] The future covariates are obtained by using the audio frequency-aging correlation. The neural network model based on the multilayer perceptron is trained according to the future covariates, the historical industrial park operation data and the historical distributed energy operation data to obtain the net load prediction model of the industrial park.

[0081] At this time, the audio-aging correlation reflects the intrinsic connection between the change of audio characteristics and equipment performance during the aging process of industrial equipment over time. The future covariates of the quantile regression model are obtained by the audio-aging correlation to improve the accuracy of quantile selection. For example, in the early stage of industrial equipment aging, the power load may be relatively stable, and the corresponding quantile value range is narrow; as the aging of industrial equipment increases, the load fluctuation increases, and the range of quantile values ​​will also become wider to reflect the increase in this uncertainty.

[0082] The multilayer perceptron network architecture (MLP) has powerful function approximation capabilities and can handle complex nonlinear relationships. In load forecasting, the relationship between power load and many factors (such as meteorological conditions, equipment operating status, time, etc.) is often nonlinear. By taking historical industrial park operation data as input, MLP can learn the potential laws between these factors and power load, thereby building a probabilistic forecasting model. For example, MLP can learn the probability distribution of power load under different meteorological conditions and equipment aging conditions, output load forecast values ​​at different probability levels, and provide probabilistic information for load forecasting.

[0083] The quantile regression model can directly quantify the uncertainty of the prediction results, and provide prediction values ​​at different quantiles by establishing the relationship between quantiles and input variables (including historical characteristics and future covariates). Since the power load is affected by many uncertain factors, such as weather variability and sudden equipment failures, the quantile regression model can accurately capture these uncertainties based on the distribution characteristics of the data, and output load prediction values ​​at different quantiles such as 0.1 quantile, 0.5 quantile (median), and 0.9 quantile.

[0084] In this embodiment, the probability prediction model can use the nonlinear processing capability of MLP to learn the complex load-factor relationship, and the quantile regression model focuses on quantifying uncertainty. It can not only take into account the comprehensive impact of various factors on the load, but also accurately give the load prediction values ​​under different confidence intervals, thereby obtaining a more accurate and reliable load forecasting result.

[0085] Building a probabilistic prediction model based on a multi-layer perceptron network architecture includes:

[0086] A probability prediction model is constructed based on the correlation between distributed energy output and load, and the correlation between environmental factors and load.

[0087] The multilayer perceptron network architecture (MLP) is a feedforward neural network consisting of multiple neuron layers, including an input layer, a hidden layer, and an output layer. The multilayer perceptron network's ability to process nonlinear relationships is used to analyze the nonlinear relationship between load demand and distributed energy output. For example, as distributed energy output increases, the relationship between load demand and net load is not a simple linear decrease, and may be affected by factors such as the charging and discharging of energy storage equipment and the operating priority of industrial equipment. MLP can capture this complex nonlinear change through the weighted summation of its neurons and the activation function operation. At the same time, the complex interaction between load demand variables and meteorological variables is also considered. For example, there may be a nonlinear correlation between temperature and load demand. When the temperature is low, the power consumption of heating equipment increases. However, when the temperature is extremely low, some industrial production processes may reduce power consumption due to low temperature restrictions. MLP can learn this complex temperature-load interaction and the synergistic effects between meteorological variables, and obtain the comprehensive effects of temperature and solar irradiance on photovoltaic power generation and load demand.

[0088] In this embodiment, a MLP-based deep neural network prediction architecture is used as a probabilistic prediction model, such as NBEATS and NHiTS.

[0089] Building a quantile regression model based on quantile regression includes:

[0090] Construct the input layer based on the rotation position embedding;

[0091] Construct a historical feature encoder and a future covariate encoder based on historical features and future covariates;

[0092] Construct feature fusion layer, quantile suggestion network and quantile value network;

[0093] A quantile regression model is constructed based on the input layer, historical feature encoder, future covariate encoder, feature fusion layer, quantile proposal network, and quantile value network.

[0094] RoPE technology uses the rotation matrix to generate an embedded representation for each data point based on the position information of the data in the time series.

[0095] The input layer is constructed to receive historical features and future covariates, and the rotated position embedding is used to generate an embedded representation. Since net load forecasting involves time series of different lengths, when processing these different length sequences, the position information is accurately encoded, so that the model can better understand the changing rules of net load data on the time axis in subsequent steps, especially when making long-term forecasts, it can distinguish the importance weights of net load data at different time points to the current forecast, thereby improving the long-term forecasting ability of the model.

[0096] For historical feature data, it is input into a special historical feature encoder. First, the historical net load data (such as two-dimensional data containing time and load values) and other historical feature data (such as meteorological data) that may be multi-dimensional are reconstructed into a one-dimensional feature representation through a flattening layer. Then, this one-dimensional feature data passes through a set of MLP blocks, Batchnorm and SiLU nonlinear activation functions in sequence. The MLP block extracts deep features from the data through weighted calculations of neurons and nonlinear transformations of activation functions; Batchnorm normalizes the input data of each layer to accelerate the training process and improve the generalization ability of the model; the SiLU function introduces nonlinearity to enhance the model's ability to fit complex relationships, and ultimately produces a feature representation of historical features.

[0097] Similarly, for future covariate data, it is input into the corresponding future covariate encoder. After similar flattening layer, MLP block, Batchnorm and SiLU function processing, the feature representation of the future covariate is generated.

[0098] By constructing a historical feature encoder and a future covariate encoder, the model can effectively extract and transform historical information and future prediction information respectively, convert the original data into a feature form that is more suitable for subsequent quantile prediction, improve the prediction accuracy of net load, and provide an accurate basis for power dispatch.

[0099] Historical characteristics refer to past data, including historical load demand , historical covariates and other influencing variables, such as historical characteristics and future covariates As an example, use nonparametric quantile regression to model the conditional distribution:

[0100] ;

[0101] Among them, p represents the probability distribution, y H is the load in the next H hours, x L is the historical feature in the past L hours, c H is the future covariate in the next H hours, is a The quantile probability set satisfies ; yes Quantile function of , where N is the number of samples.

[0102] Quantiles and quantile functions are represented by two feed-forward networks (FFNs): a quantile proposal network (FPN) and a quantile value network (QVN). The FPN learns to generate quantiles, while the QVN maps these quantiles to quantile functions, enabling end-to-end learning of quantiles and quantile functions.

[0103] Specifically, the feature representations generated by the historical feature encoder and the future covariate encoder are input to the quantile proposal network (FPN). FPN is a feed-forward network (FFN) with a dropout layer. The dropout layer randomly drops neuron connections with a certain probability during training to prevent overfitting of the model and improve its generalization ability. FPN determines the appropriate quantile value based on the input feature representation through calculation and parameter adjustment of internal neurons. The output is converted into quantile ratios using the Softmax activation function to ensure that these ratios meet the requirements of the probability distribution (the sum is 1). When faced with different historical net load trends and changes in meteorological forecasts, FPN can determine which quantiles are more critical to describing the uncertainty of the net load, thereby generating a corresponding set of quantiles, which provides a basis for the subsequent calculation of quantile values.

[0104] The quantile value network (QVN) receives the quantile ratio generated by FPN and the feature representation generated by the encoder. QVN captures the interaction between the quantile ratio and the feature representation, and determines the specific value range of the net load at different quantiles through its own network structure and parameter calculation. Based on this information and the input data, QVN calculates the specific predicted value of the net load at a specific quantile. Through this step, the conversion from the original input data to the net load quantile estimate is realized, and finally the probabilistic prediction of the net load is completed, and the non-parametric probability density estimate is obtained, and the uncertainty of the prediction is quantified with a confidence interval.

[0105] In this embodiment, constructing a neural network model based on a multilayer perceptron according to the probability prediction model and the quantile regression model includes:

[0106] A neural network model based on multi-layer perceptron is constructed by fusing the encoders of the probability prediction model and the quantile regression model.

[0107] Probabilistic prediction models are usually modeled based on certain assumptions and distribution forms, and may be more sensitive to changes in data distribution. Quantile regression models are less dependent on specific distribution assumptions and are more adaptable to data. Fusion of the encoders of the two makes the model more robust in the face of changes in data distribution and reduces the risk of overfitting.

[0108] In this embodiment, the future covariates are obtained by using the audio-aging correlation, and the net load prediction model of the industrial park is obtained by training the neural network model based on the multilayer perceptron according to the future covariates, the historical industrial park operation data, and the historical distributed energy operation data. The future covariates are obtained by using the audio-aging correlation, and the equipment aging factor can be incorporated into the net load prediction model, providing a comprehensive capture of various factors affecting the net load and their interrelationships. The future covariates obtained based on the audio-aging correlation take into account the impact of the future state changes of the equipment on the net load, so as to better adapt to changes in equipment updates, process improvements, etc. in the industrial park.

[0109] Among them, the neural network model based on the multilayer perceptron is constructed according to the probability prediction model and the quantile regression model, and also includes:

[0110] Construct quantile loss function and constraints based on Huber function and 1-Wasserstein metric;

[0111] A neural network model based on multilayer perceptron is constructed based on the probability prediction model, the quantile regression model quantile loss function and constraints.

[0112] The uncertainty interval in the prediction is linearized by the quantile loss function and constraint conditions, and the prediction error is smoothed to ensure that the model can maintain robust prediction capabilities when facing complex weather and load fluctuations.

[0113] Construct the initial loss function by minimizing the pinball loss function:

[0114] ;

[0115] in, is the error term, representing the nth observation at time t in the quantile The residual under is a value between 0 and 1, indicating the quantile that the model wants to fit; T is the total number of time steps, t is the time, and N is the number of observations.

[0116] Since the pinball loss has a non-differentiable point at the origin, the Huber function is used for smoothing approximation to obtain the Huber quantile loss:

[0117] ;

[0118] in, represents the indicator function, when hour, The value of is 1, otherwise it is 0, which is used to distinguish the positive and negative values ​​of the prediction error; Adjust the parameters for the loss, About Huber loss function.

[0119] To further optimize the accuracy of the predicted quantiles, a penalty term is added:

[0120] ;

[0121] Where: is the discount factor, ; Represents the trajectory of the strategy, and the input is the quantile probability set at time t ; is a scalar value that controls the size of the penalty term, is the quantile probability set of the n+1th observation at time t, is the quantile probability set of the nth observation at time t.

[0122] The 1-Wasserstein metric is incorporated into the learning process of the quantile score to obtain the quantile loss function.

[0123] The constraints include at least: parameter adjustment constraints in Huber quantile loss, parameter control constraints in penalty terms, and 1-Wasserstein metric related constraints.

[0124] Among them, the parameter adjustment in the Huber quantile loss controls the sensitivity of the model to small errors by changing the parameter value, thereby affecting the quantile prediction effect; the parameter control in the penalty term includes controlling the size of the penalty term through parameters to make the model adjust the parameters cautiously, and adjusting the quantile sharpness through parameters to flexibly control the degree of difference in predicted values ​​under different quantiles; 1-Wasserstein metric-related constraints constrain model parameter adjustment through their definition and derivative regulations, so that the predicted quantile distribution is closer to the true distribution, thereby improving the prediction quality and reliability.

[0125] In this embodiment, step S4 further includes:

[0126] Acquire the current industrial park operation data, and filter out the human voice from the current industrial park audio data in the current industrial park operation data;

[0127] The fluctuation parameters within a preset time period are obtained by using the current industrial park audio data after human voice screening and the audio-aging correlation;

[0128] Output the distributed energy output forecast data within a preset period according to the current distributed energy operation data and the distributed energy output forecast model;

[0129] The distributed energy output forecast data, environmental data within a preset period of time, and fluctuation parameters within a preset period of time are used as future covariates of the industrial park load net forecast model to output the industrial park net load forecast data.

[0130] Among them, the preset time period is the target prediction time period, and the distributed energy output prediction model can be a photovoltaic prediction model, such as a multi-site photovoltaic power generation prediction model based on a graph attention network, or it can be any other neural network model that can predict future distributed energy output.

[0131] In some other cases, step S4 further includes:

[0132] Construct a periodic fluctuation curve based on the historical repetition cycle, and construct a quantile constraint interval based on the periodic fluctuation curve;

[0133] The distributed energy output forecast data, environmental data within a preset period and fluctuation parameters within a preset period are used as future covariates of the industrial park load net forecast model. The quantile of the industrial park load net forecast model is selected according to the quantile constraint interval to output the industrial park net load forecast data.

[0134] In this case, since the audio frequency repetition period may change, the period fluctuation of the future time series is predicted based on the changes in the historical repetition period, such as the continuous reduction or increase of the period time, so as to construct the quantile constraint interval. The quantile constraint interval refers to the upper and lower quantiles within a certain probability range (such as a 90% confidence interval) in the probability distribution. For example, if it is predicted that the audio frequency repetition period may fluctuate within a range in a certain period of time in the future, then the range of values ​​of the quantiles in the net load prediction model can be determined based on this fluctuation range. The quantile selection of the net load prediction model of the industrial park is controlled so that the confidence interval contains the uncertainty brought by the audio frequency repetition period, further improving the accuracy of the net load prediction data of the industrial park.

[0135] Step S4 also includes:

[0136] The automatic regression reasoning mechanism is used to perform rolling prediction of the net load data of the industrial park according to the preset time period based on the current industrial park operation data, the current distributed energy operation data, the industrial park load net prediction model and the sound frequency-aging correlation to obtain the net load prediction data of the industrial park.

[0137] Rolling forecasts are performed according to preset time periods. For example, the preset time period can be every 15 minutes, every hour, or every day, etc., depending on the power dispatching needs and data collection frequency of the industrial park. At each forecast moment, the ARI mechanism forecasts the net load based on the current existing data (current industrial park operation data, current distributed energy operation data) and the industrial park load net forecast model. Then, as time goes by, in the next preset time period, the newly acquired data is added to the forecast process and the forecast is re-performed, and this is continuously rolled out to ensure the timeliness of the output of the net load forecast data of the industrial park.

[0138] As a third embodiment of the present application, an industrial park net load prediction system based on a multi-layer perceptron includes:

[0139] The audio processing module builds audio-aging correlations based on historical industrial park operation data;

[0140] Model building module, building a neural network model based on a multi-layer perceptron, and obtaining an industrial park net load prediction model based on the sound frequency-aging correlation and neural network model training;

[0141] The fusion analysis module outputs the net load forecast data of the industrial park based on the current industrial park operation data and the current distributed energy operation data.

[0142] In this embodiment, the audio processing module divides the historical industrial park operation data according to the audio collection area, obtains the regional historical operation data corresponding to the audio collection area, performs cluster analysis on the historical operation data of each area based on the audio fluctuation, and obtains the audio-aging correlation with the clustering result. The fusion analysis module is connected to the audio processing module and the model building module, and uses the historical industrial park operation data, the historical distributed energy operation data and the audio-aging correlation to train the neural network model based on the multilayer perceptron to obtain the industrial park net load prediction model, obtain the current industrial park operation data and the current distributed energy operation data, and obtain the industrial park net load prediction data with the current industrial park operation data, the current distributed energy operation data, the industrial park load net prediction model and the audio-aging correlation.

[0143] In some cases, a distributed collaborative approach can be adopted, that is, an audio processing module is set for each producer, so that the audio processing results are uploaded by the producer, which improves the privacy security of the producer and improves the data processing efficiency. At the same time, the fusion analysis module is also connected to the distributed energy output prediction model to obtain distributed energy output prediction data. It can also be that the model construction module has a distributed energy output prediction model pre-built. The fusion analysis module calls the audio-aging correlation output of the audio processing module according to the current industrial park operation data and the current distributed energy operation data, the distributed energy output prediction model constructed by the model construction module, and the industrial park net load prediction model to output the industrial park net load prediction data.

[0144] The specific implementation described above is a preferred implementation of the industrial park net load prediction method and system based on a multi-layer perceptron of the present application, and is not intended to limit the specific implementation scope of the present application. The scope of the present application includes but is not limited to the specific implementation embodiment. All equivalent changes made in accordance with the shape and structure of the present application are within the protection scope of the present application.

Claims

1. The net load forecasting method of industrial park based on multi-layer perceptron is characterized by: The steps include: S1: Acquire historical industrial park operation data, divide the historical industrial park operation data according to the audio collection area, and obtain regional historical operation data corresponding to the audio collection area; S2: Perform cluster analysis on the historical operation data of each region based on the sound frequency fluctuation, and use the clustering results to obtain the sound frequency-aging correlation; S3: Using historical industrial park operation data, historical distributed energy operation data and sound frequency-aging correlation to train a neural network model based on a multi-layer perceptron to obtain an industrial park net load prediction model; S4: obtaining the current industrial park operation data and the current distributed energy operation data, and obtaining the industrial park net load prediction data by using the current industrial park operation data, the current distributed energy operation data, the industrial park load net prediction model and the sound frequency-aging correlation; Screening out human voices from the audio data of the current industrial park in the operation data of the current industrial park; The fluctuation parameters within a preset time period are obtained by using the current industrial park audio data after human voice screening and the audio-aging correlation; Construct a periodic fluctuation curve based on the historical repetition cycle, and construct a quantile constraint interval based on the periodic fluctuation curve; The distributed energy output forecast data, environmental data within a preset period and fluctuation parameters within a preset period are used as future covariates of the industrial park load net forecast model. The quantile of the industrial park load net forecast model is selected according to the quantile constraint interval to output the industrial park net load forecast data.

2. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 1, characterized in that: The S1 further comprises: Acquire historical industrial park operation data, perform analog-to-digital conversion on the historical industrial park audio data, and perform human voice screening on the analog-to-digital converted historical industrial park audio data; The historical industrial park operation data after human voice screening is divided according to the audio collection area to obtain the regional historical operation data corresponding to the audio collection area.

3. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 2, characterized in that: The human voice screening process of the analog-to-digital converted historical industrial park audio data also includes: Analyze the frequency distribution of the historical industrial park audio data after analog-to-digital conversion using spectrum analysis, and filter out human voices based on the frequency band differences between the sounds of industrial equipment operation and human voices; According to the historical industrial park audio data after the first human voice screening, a second human voice screening is performed based on the repetition number threshold and the repetition cycle threshold.

4. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 1, characterized in that: The S2 further includes: Extract features from historical operation data of each region to obtain audio frequency fluctuation features; Clustering is performed according to the audio frequency fluctuation characteristics by using a clustering algorithm to obtain a clustering result corresponding to the audio frequency fluctuation characteristics; For each clustering result, the continuous operation time of the industrial equipment in the clustering result is calculated, and the audio frequency-aging correlation is constructed according to the time series relationship of the audio frequency fluctuation characteristics and the continuous operation time.

5. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 4, characterized in that: The step of constructing the audio frequency-aging correlation according to the temporal relationship of the audio frequency fluctuation characteristics and the continuous operation time includes: According to the corresponding relationship between the audio frequency fluctuation characteristics and the continuous running time, the audio frequency fluctuation characteristic smoothing weight is obtained; The mutation weight of the acoustic frequency fluctuation feature is obtained according to the continuous operation time of the industrial equipment in each clustering result; The audio frequency-aging correlation is obtained based on the smooth weight of the audio frequency fluctuation feature and the sudden weight of the audio frequency fluctuation feature.

6. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 5, characterized in that: The S3 includes: According to the historical industrial park audio data and historical time series, the fluctuation parameters are constructed according to the smooth weight of the audio fluctuation characteristics and the sudden weight of the audio fluctuation characteristics; The net load prediction model of the industrial park is obtained by training the neural network model based on the multi-layer perceptron using the historical industrial park operation data, historical distributed energy operation data and fluctuation parameters.

7. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 1, characterized in that: The S3 further includes: Construct a probabilistic prediction model based on a multi-layer perceptron network architecture; Construct a quantile regression model based on quantile regression; Construct a neural network model based on multilayer perceptron according to the probability prediction model and quantile regression model; The future covariates are obtained by using the audio frequency-aging correlation. The neural network model based on the multilayer perceptron is trained according to the future covariates, the historical industrial park operation data and the historical distributed energy operation data to obtain the net load prediction model of the industrial park.

8. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 7, characterized in that: The probability prediction model constructed based on the multi-layer perceptron network architecture includes: A probability prediction model is constructed based on the correlation between distributed energy output and load, and the correlation between environmental factors and load.

9. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 7, characterized in that: The constructing of a quantile regression model based on quantile regression includes: Construct the input layer based on the rotation position embedding; Construct a historical feature encoder and a future covariate encoder based on historical features and future covariates; Construct feature fusion layer, quantile suggestion network and quantile value network; A quantile regression model is constructed based on the input layer, historical feature encoder, future covariate encoder, feature fusion layer, quantile proposal network, and quantile value network.

10. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 7, characterized in that: The neural network model based on the multilayer perceptron is constructed according to the probability prediction model and the quantile regression model, and further includes: Construct quantile loss function and constraints based on Huber function and 1-Wasserstein metric; A neural network model based on multilayer perceptron is constructed based on the probability prediction model, the quantile regression model quantile loss function and constraints.

11. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 7, characterized in that: S4 also includes: Acquire the current industrial park operation data, and filter out the human voice from the current industrial park audio data in the current industrial park operation data; The fluctuation parameters within a preset time period are obtained by using the current industrial park audio data after human voice screening and the audio-aging correlation; Output the distributed energy output forecast data within a preset period according to the current distributed energy operation data and the distributed energy output forecast model; The distributed energy output forecast data, environmental data within a preset period of time, and fluctuation parameters within a preset period of time are used as future covariates of the industrial park load net forecast model to output the industrial park net load forecast data.

12. The method for predicting net load of an industrial park based on a multi-layer perceptron as claimed in claim 1, characterized in that: The S4 further comprises: The automatic regression reasoning mechanism is used to perform rolling prediction of the net load data of the industrial park according to the preset time period based on the current industrial park operation data, the current distributed energy operation data, the industrial park load net prediction model and the sound frequency-aging correlation to obtain the net load prediction data of the industrial park.

13. A multi-layer perceptron-based industrial park net load forecasting system, used to implement the method according to any one of claims 1 to 12, characterized in that: include: The audio processing module builds audio-aging correlations based on historical industrial park operation data; Model building module, building a neural network model based on a multi-layer perceptron, and obtaining an industrial park net load prediction model based on the sound frequency-aging correlation and neural network model training; The fusion analysis module outputs the net load forecast data of the industrial park based on the current industrial park operation data and the current distributed energy operation data.

Images