Park heterogeneous load prediction method based on EEMD and Kmeans adaptive optimization

Abstract

The invention discloses a park heterogeneous load prediction method based on EEMD and Kmeans adaptive optimization. The method comprises the following steps: carrying out missing value filling, abnormal value elimination and feature coding on park historical load data and external influence factors; dividing the park load into a plurality of typical scenes through a Kmeans clustering algorithm based on the similarity of load fluctuation characteristics; eEMD decomposition is carried out on the load sequence corresponding to each typical scene, Gaussian white noise is added to suppress modal aliasing, multi-frequency components are extracted and fused with external variables, and a prediction input data set is constructed; a prediction model is selected for different scenes, and adaptive adjustment and optimization are carried out through a Bayesian optimization algorithm; and carrying out dynamic weight fusion on the prediction result of each scene, dynamically adjusting the weight according to the historical load proportion of the scene, and outputting an overall load prediction value of the park.

Description

Technical Field

[0001] The present invention relates to the field of energy management technology, and in particular to a park heterogeneous load forecasting method based on EEMD and Kmeans adaptive optimization. Background Art

[0002] Electricity is a core resource for the development and daily operations of modern industrial parks. Accurate industrial park power load forecasting is crucial to ensuring the stability of the park's power supply. The power load within the industrial park is heterogeneous, meaning it accounts for different types of electricity demand and fluctuating characteristics, including various electricity consumption modes such as industrial and commercial. By accurately forecasting industrial park power demand, power supply companies can achieve efficient grid load planning and optimize power generation and power dispatch within the park, thereby avoiding power shortages or surpluses during peak load periods, reducing operating costs, and improving the efficiency of the power system. Furthermore, accurate load forecasting can help power companies reduce energy waste caused by excessive power generation, promote green and low-carbon development of industrial parks, and reduce greenhouse gas emissions such as carbon dioxide.

[0003] Industrial park loads are typically highly heterogeneous, encompassing multiple sectors such as industry, commerce, and offices, with each load type exhibiting significantly different fluctuation characteristics. Furthermore, loads within industrial parks are not only affected by external factors such as seasonality and weather, but may also be disrupted by complex internal and external factors such as real-time electricity price fluctuations, emergencies, and policy adjustments. Traditional load forecasting methods, which mostly rely on time series analysis, ignore the combined impact of these variable factors on load fluctuations, resulting in low forecasting accuracy. To address this challenge, it is necessary to deeply explore the inherent characteristics of industrial park load data, identify potential correlations and patterns between different load types, and select targeted and adaptable forecasting models.

[0004] The application number is: CN201910294713.6, and the name is: Chinese patent for a method for screening input features of load forecasting based on hierarchical clustering and neural networks. First, the original data set including the load forecast is clustered using hierarchical clustering analysis, and the data set is divided into different classes according to its characteristics, and the features in each class are analyzed; then, the prediction effect of different features in each class is analyzed using a neural network, and finally the features with poor prediction effect are eliminated to obtain the final optimized input feature set for building load forecasting. This method uses a single model to construct a neural network, and does not take into account that there is room for further optimization at the model structure level. When considering multi-dimensional influencing factors such as meteorological factors and holidays, the temporal information contained in the load time series itself is not fully explored, resulting in the failure to effectively capture the inherent laws and correlations in the load data, and thus the inability to better capture the relevant factors in load forecasting.

[0005] The application number is CN202410128279.5, and the title is "A Chinese patent for a load forecasting method considering multi-scale time series information." Based on historical load data and Prophet load time series components, the Pearson correlation coefficient method is used for correlation analysis to construct a data feature set. A wide expansion coefficient LSTM-TCN multi-scale time series mining module is constructed and trained to simultaneously capture and learn short-term time series information at the hourly level and time series information at the cycle level. The output of the trained LSTM-TCN time series mining module is then used as the input data for the basic decision tree model of the GBRT algorithm, and the information after multi-scale time series processing is used as the output data for training the basic decision tree model. The final basic decision tree model is obtained to enhance the robustness of the model prediction and achieve load forecasting that considers multi-scale time series information. This method considers the time series information within the multi-scale load time series and is suitable for scenarios with periodic regularities and smooth and stable demand response. However, relying solely on this method to build short-term load forecasting models for non-stationary signals, which are significantly influenced by temporal and spatial factors, has certain limitations. Specifically, this method cannot effectively process and separate the complex components of non-stationary signals, nor can it fully capture the multi-scale and multi-frequency characteristics of the signals. Therefore, there is an urgent need to introduce more adaptive forecasting and analysis methods that can decompose non-stationary signals into components of different scales and frequencies to better understand and analyze the inherent characteristics of load series. Summary of the Invention

[0006] In response to the defects and shortcomings of existing technologies, the present invention provides a campus load forecasting method and system based on heterogeneous load classification and multi-scale signal decomposition. The system addresses the limitations of traditional methods in non-stationary signal processing, scenario diversity modeling, and dynamic optimization through the following innovative designs:

[0007] Heterogeneous load classification and decomposition collaborative mechanism: The Kmeans clustering algorithm is used to accurately classify heterogeneous load scenarios such as industrial and commercial loads in the park, and integrated empirical mode decomposition (EEMD) is performed on the load sequence of each scenario. Modal aliasing is suppressed through noise injection and multiple decomposition mean calculations, and multi-frequency components (IMFs) adapted to the characteristics of the scenario are extracted, overcoming the shortcomings of traditional methods (such as hierarchical clustering or fixed-period decomposition) in processing non-stationary signals.

[0008] Scenario-adaptive modeling and dynamic optimization: Differentiated prediction models (edRVFL, BiLSTM, SVR) are selected for different load scenarios, and hyperparameters are dynamically adjusted through the Bayesian optimization algorithm (TPE strategy) to solve the problems of poor generalization ability of a single model and low efficiency of manual parameter adjustment.

[0009] Closed-loop full-link design: From data preprocessing (missing value filling, outlier removal), feature fusion (meteorological and holiday variable encoding) to dynamic weight fusion of prediction results, an end-to-end process of classification → decomposition → optimization → fusion is formed, which significantly improves prediction accuracy (experimental data shows that RMSE is reduced by more than 30%) and adapts to the real-time, heterogeneous and non-stationary requirements of park loads under power market reform.

[0010] This solution provides a high-precision and high-robustness solution for complex load forecasting in smart grids through the synergistic effect of classification and decomposition, scenario adaptation modeling and closed-loop optimization process.

[0011] The present invention specifically adopts the following technical solutions:

[0012] A method for predicting heterogeneous loads in a park based on EEMD and Kmeans adaptive optimization includes the following steps:

[0013] Fill missing values, remove outliers and perform feature coding on the park's historical load data and external influencing factors;

[0014] Based on the similarity of load fluctuation characteristics, the park load is divided into multiple typical scenarios using the Kmeans clustering algorithm;

[0015] Perform EEMD decomposition on the load sequence corresponding to each typical scenario, suppress modal aliasing by adding Gaussian white noise, extract multi-frequency components and fuse them with external variables to construct a prediction input data set;

[0016] Select prediction models for different scenarios and adaptively tune them using the Bayesian optimization algorithm;

[0017] The prediction results of each scenario are dynamically weighted and fused, and the weight is dynamically adjusted according to the historical load ratio of the scenario to output the overall load prediction value of the park.

[0018] Furthermore, the implementation of the Kmeans clustering specifically includes:

[0019] The number of clusters is preset based on the park load type, and the initial cluster center is dynamically adjusted according to the historical load fluctuation curve;

[0020] The scenario labels generated by clustering are directly used as the input of EEMD decomposition to ensure that the decomposition process only targets the same type of load fluctuation characteristics;

[0021] The cluster centers are updated iteratively until the Euclidean distance variance of the load series within the cluster is less than the threshold.

[0022] Furthermore, the implementation of the EEMD decomposition specifically includes:

[0023] The load sequence of each scenario is decomposed multiple times independently. Before each decomposition, Gaussian white noise of different instances is superimposed. The noise amplitude is inversely proportional to the square root of the number of decompositions.

[0024] The decomposition process is dynamically bound to the classification results to ensure that the load fluctuation characteristics of industrial and commercial scenarios are independently characterized by differentiated frequency band components.

[0025] Furthermore, the Bayesian optimization adopts a tree-based Parzen window estimation strategy, the objective function is the root mean square error of the validation set, and the hyperparameters are selected through the expected improvement function.

[0026] Furthermore, the weight of the dynamic weight fusion is calculated based on the historical load ratio of each scenario.

[0027] Furthermore, the prediction models include edRVFL, BiLSTM, and SVR, and the model type is selected based on the load fluctuation characteristics of the scenario:

[0028] The edRVFL model is selected for industrial load scenarios;

[0029] The BiLSTM model is selected for commercial load scenarios;

[0030] Select the SVR model for optical cable load scenarios.

[0031] Furthermore, the external influencing factors include weather conditions, holiday types, and real-time electricity prices;

[0032] The feature coding includes:

[0033] Map weather conditions to numeric labels;

[0034] The holiday type, day of the week, and time point are expanded to the Euclidean space through one-hot encoding.

[0035] And, a campus heterogeneous load forecasting system based on EEMD and Kmeans adaptive optimization, including:

[0036] Data preprocessing module: This module is configured to fill missing values, remove outliers, and perform feature encoding on the park's historical load data and external influencing factors, and output a standardized data set;

[0037] Heterogeneous load classification module: This module is configured to use the Kmeans clustering algorithm to divide the campus load into multiple typical scenarios based on the similarity of load fluctuation characteristics, and output scenario labels and classified load sequences.

[0038] Multi-scale signal decomposition module: This module is configured to perform EEMD decomposition on the load sequence of each scenario, suppress modal aliasing by adding Gaussian white noise, extract multi-frequency components and residual components, and fuse them with external variables to generate a multi-dimensional feature dataset;

[0039] Scenario Adaptive Modeling and Optimization Module: This module is configured to select prediction models for different scenarios, adaptively tune model hyperparameters using the Bayesian optimization algorithm, and output the trained prediction model.

[0040] Prediction result fusion module: It is configured to dynamically weight the prediction results of each scenario. The weight is calculated based on the historical load ratio of the scenario, and the overall load prediction value of the park is output.

[0041] Furthermore, the heterogeneous load classification module includes:

[0042] Dynamic initialization unit: sets the initial number of clusters based on the preset park load type and dynamically adjusts the cluster center according to the historical load fluctuation curve;

[0043] Iterative optimization unit: through Euclidean distance calculation and cluster center update, until the Euclidean distance variance of the load sequence within the cluster is less than the preset threshold;

[0044] Classification-decomposition interface unit: Directly transmits the scene labels generated by clustering to the multi-scale signal decomposition module to ensure that the decomposition process is only for the same type of load data;

[0045] The multi-scale signal decomposition module includes:

[0046] Noise injection unit: superimposes Gaussian white noise on the classified load sequence, and the noise amplitude is inversely proportional to the square root of the decomposition number;

[0047] Modal aliasing suppression unit: eliminates noise interference and extracts stable IMF components through multiple independent decompositions and result mean calculations;

[0048] Feature binding unit: aligns the decomposed multi-frequency components with meteorological and holiday features by timestamp to construct a spatiotemporally correlated input dataset.

[0049] Furthermore, the scenario adaptive modeling and optimization module includes:

[0050] Model selection unit: Assigns prediction models based on the load fluctuation characteristics of the scenario. The edRVFL model is used for industrial load scenarios, the BiLSTM model is used for commercial load scenarios, and the SVR model is used for optical cable load scenarios.

[0051] Bayesian optimization unit: uses the tree-based Parzen window estimation algorithm, takes the root mean square error of the validation set as the objective function, and searches for the optimal solution of hyperparameters through the expected improvement function;

[0052] The prediction result fusion module includes:

[0053] Weight calculation unit: dynamically adjusts weights based on the historical load ratio of the scenario;

[0054] Fusion execution unit: performs weighted summation of the prediction values of each scene and iteratively optimizes the weight distribution through the error reverse correction mechanism.

[0055] The system is deployed on a distributed computing platform, including:

[0056] Data collection layer: Real-time acquisition of park load, weather, and electricity price data through IoT sensors;

[0057] Edge computing nodes: perform data preprocessing and heterogeneous load classification to reduce cloud transmission latency;

[0058] Cloud computing center: runs multi-scale signal decomposition, model optimization and result fusion algorithms to provide high-precision load forecasting services.

[0059] And, an electronic device includes a memory, a processor, and a computer program stored in the memory and runnable on the processor, characterized in that the processor implements the steps of the above method when executing the program.

[0060] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0061] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0062] Improve non-stationary signal processing capabilities: Through the synergistic mechanism of Kmeans clustering and EEMD decomposition, it effectively separates non-stationary signals in heterogeneous load scenarios, suppresses modal aliasing, and significantly improves the stability and accuracy of multi-frequency feature extraction, overcoming the shortcomings of traditional methods in their lack of adaptability to complex time series signals.

[0063] Enhance scenario-differentiated modeling capabilities: Differentiate prediction models (such as edRVFL, BiLSTM, SVR) are selected for different load scenarios such as industrial and commercial scenarios, and hyperparameters are dynamically adapted through Bayesian optimization to improve the model's ability to capture diverse load fluctuation characteristics and avoid the generalization limitations of a single model.

[0064] Optimize prediction efficiency and automation level: Use the Bayesian optimization algorithm (TPE strategy) to replace manual parameter adjustment or grid search, significantly reducing the time consumption of hyperparameter optimization. At the same time, adaptively integrate multi-scenario prediction results through dynamic weight fusion strategy to reduce manual intervention.

[0065] The full-link closed-loop design improves robustness: from data preprocessing (missing value filling, outlier removal), feature fusion (meteorological, holiday day coding) to result output, an end-to-end load forecasting process is formed, which systematically solves the error accumulation problem caused by the traditional segmented processing method and adapts to the dynamic impact of complex external factors (such as real-time electricity prices and policy disturbances).

[0066] Through the above-mentioned technical means, this solution achieves efficient and adaptive modeling of heterogeneous load scenarios in the park while ensuring prediction accuracy, providing reliable support for the refined operation of the smart grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0068] Figure 1 This is a diagram of the architecture of the park heterogeneous load forecasting process model based on EEMD and Kmeans adaptive optimization in an embodiment of the present invention;

[0069] Figure 2 This is a flow chart of a load forecasting model under different typical scenarios according to an embodiment of the present invention;

[0070] Figure 3 This is a schematic diagram of the box plot outlier detection principle of an embodiment of the present invention;

[0071] Figure 4 This is a diagram of the Kmeans clustering principle of an embodiment of the present invention;

[0072] Figure 5 This is a flow chart of the EMD algorithm according to an embodiment of the present invention;

[0073] Figure 6 This is a flowchart of the hyperparameter optimization process based on Bayesian optimization in an embodiment of the present invention. DETAILED DESCRIPTION

[0074] Hereinafter, specific embodiments of the present application will be described in detail with reference to the accompanying drawings. Based on these detailed descriptions, those skilled in the art will be able to clearly understand the present application and implement the present application. Without violating the principles of the present application, the features of different embodiments may be combined to obtain new implementations, or certain features of certain embodiments may be substituted to obtain other preferred implementations.

[0075] To make the features and advantages of the present invention more clearly understood, the following embodiments are specifically described in detail with reference to the accompanying drawings.

[0076] Against the backdrop of power market reform and smart grid construction, industrial park power load forecasting faces complex challenges. Traditional time series analysis methods ignore the combined impact of multiple factors, resulting in low forecasting accuracy. The proposed method reduces the complexity of load series through Kmeans clustering to achieve accurate classification modeling. It uses EEMD decomposition to capture the signal characteristics of each frequency band and reconstructs the input dataset by combining multidimensional time series variables. It constructs forecasting models and optimizes hyperparameters for different scenarios, effectively addressing the accuracy bottleneck of traditional methods that struggle to cope with the multi-factor coupling impact of heterogeneous loads in industrial parks, providing support for accurate forecasting, supervision, and sustainable development of power grid loads.

[0077] The embodiment of the present invention is implemented through the following steps: the data acquisition module acquires the electricity consumption data of multiple types of loads in the park in real time; the load feature analysis module uses the Kmeans clustering algorithm to perform in-depth scene division on heterogeneous loads to generate load category clusters with similar fluctuation characteristics; the signal processing module uses the EEMD algorithm to decompose various load sequences into K IMFs components and RES residual components, and performs feature fusion with multidimensional time series variables such as meteorology and electricity prices; the model optimization module constructs edRVFL, BiLSTM and SVR prediction models for different load categories, and uses the Bayesian optimization algorithm to automatically optimize hyperparameters; the prediction execution module weightedly fuses the prediction results of each component and outputs the overall load prediction curve of the park.

[0078] Please refer to the following steps for details:

[0079] Step 1: Determine the historical load series, meteorological factors, holiday types, time points, and other factors collected by the data acquisition system as sample characteristics. Preprocess missing values using the Lagrange interpolation method and select and eliminate outliers. Perform one-hot encoding on other influencing factors besides load.

[0080] Step 1.1 Use the Lagrange interpolation method to fill in the missing values in the load data and use the box plot to identify outliers.

[0081] Step 1.1.1: Using n+1 discrete points The interpolation polynomial L can be obtained by the interpolation function n (x), as shown in formulas (1-1) and (1-2).

[0082]

[0083]

[0084] Where L n (x) is the Lagrange interpolation function; l i (x) is the Lagrange interpolation basis function.

[0085] Use the known numbers to find the Lagrange interpolation basis function, then substitute the node to be found into the polynomial to find the target value, and complete the filling of the blank values in the load data.

[0086] Step 1.1.2: Use box plots to identify outliers. Outliers are defined as those with values greater than Q. U +1.5IQR or less than Q L -1.5IQR value. U is the upper quartile, indicating that among all observations The data is bigger than his, Q L is the lower quartile, indicating that The data is smaller than his. IQR is the interquartile range, which is Q U and Q L The difference between the two contains half of the observed value, and the corresponding principle is as follows Figure 3 shown.

[0087] Step 1.1.3: Data normalization: Use the linear function normalization method to make the original data dimensionless, as shown in formula (1-3).

[0088]

[0089] Among them, x is the original data, x min and x max are the minimum and maximum values of the original data, respectively, x nor It is the standard data with the interval range of [0,1] after normalization.

[0090] Step 1.2 considers meteorological factors, including weather conditions (e.g., sunny is weather 0, cloudy is weather 1, light rain is weather 2, moderate rain is weather 3, and heavy rain is weather 4), maximum temperature, minimum temperature, relative humidity, wind speed and direction (e.g., northeast wind is 1, southeast is 2, northwest wind is 3, southwest wind is 4, north wind is 5, south wind is 6, west wind is 7, and east wind is 8); for holiday factors, decompose and extract features such as "week number" (e.g., Monday is 1, ..., Sunday is 7), "whether it is a weekend" (e.g., weekday is 0, holiday is 1), and convert them into recognizable digital signals.

[0091] One-hot encoding is used to expand the values of discrete features to Euclidean space. A certain value of the discrete feature corresponds to a certain point in Euclidean space, making the distance calculation between features more reasonable.

[0092] Step 2: Use K-means clustering algorithm to cluster the pre-processed park heterogeneous load time series Perform cluster analysis and divide the load data into multiple typical load scenarios based on their similarity to obtain the clustered load sequence. The corresponding cluster labels are The corresponding principle is as follows Figure 4 shown.

[0093] Step 2.1 Randomly initialize the cluster center. For a given set of data Where n is the number of samples of the data, and m is the dimension of the data. Initialize the cluster centers to μ1, μ2,…, μ N , where N is the number of clusters.

[0094] Step 2.2 calculates the Euclidean distance of each data point to the cluster center and assigns the data point to the cluster closest to it. i To each cluster center μ q The Euclidean distance is calculated and the cluster center with the closest distance is selected. The Euclidean distance formula is:

[0095]

[0096] Among them, x ij For data x i The component in the jth dimension, μ qj is the cluster center μ q ,q=1,2,…,N is the component in the j-th dimension.

[0097] Then, for each data x i Classify it into the cluster closest to it q=1,2,…,N, that is:

[0098]

[0099] Step 2.3: Recalculate the cluster center based on the obtained cluster. The new cluster center is the mean of all data points in the cluster:

[0100]

[0101] in, It is a cluster The number of data points in .

[0102] Step 2.4 iteratively updates until the cluster center no longer changes or the change is less than the specified threshold. Repeat steps 2.2 and 2.3 until the change in the cluster center is less than the set threshold ε, that is:

[0103]

[0104] in, is the Euclidean distance of cluster center change, and ε is the preset convergence threshold.

[0105] Step 3: Use EMD algorithm to analyze typical scenes The power load time series is stabilized to obtain the eigenmode for prediction. The corresponding principle is as follows Figure 5 shown.

[0106] Step 3.1 Apply the EMD algorithm to perform empirical mode decomposition on the load time series and obtain K IMF components i ,i=1,2,…,k and a residual component RES.

[0107] Step 3.1.1: Use the power load time series in step 2 as the initial input sequence and perform the following process on the initial sequence until the stopping criterion is met:

[0108] (a) Initialize r0=x(t), i=1, h0=r i-1 (t), j = 1, where x(t) is the original sequence signal and h0 is the initial eigenmode component.

[0109] (b) By obtaining the sequence h j-1 The local extreme points in the sequence are extracted by the maximum and minimum values of (t) to form the upper and lower envelopes.

[0110] (c) Calculate the average value of the upper and lower envelopes to obtain the mean function of the current sequence. The mean function formula is shown in formula (3-1):

[0111]

[0112] Among them, m j-1 (t) is the mean function of the j-1th iteration, up j-1 (t) is the upper envelope of the current iteration, low j-1 (t) is the lower envelope of the current iteration.

[0113] (d). From the original sequence h j-1 (t) Subtract the j-1th mean function to obtain the function h containing the high-frequency signal part j (t), h j (t) The formula is shown in formula (3-2):

[0114] h j (t) = h j-1 (t)-m j-1 (t) (3-2)

[0115] (e). Determine the current h j (t) Whether it is a qualified IMF. A qualified IMF must meet two conditions:

[0116] The number of extreme points and the number of zero crossings in the sequence are equal or differ by no more than 1.

[0117] Over the entire series, the average value of the amplitude at the extreme points is zero.

[0118] (f) If the currently obtained IMF does not meet the conditions of a qualified IMF, it is used as a new initial sequence, the loop condition j=j+1 is updated, and step 3.1.1-(b) is repeated.

[0119] (g) If the currently obtained IMF meets the conditions for a qualified IMF, it is saved, that is, the formula is satisfied:

[0120] IMF i (t) = h j (t) (3-3)

[0121] And subtract the IMF from the initial sequence, that is, satisfy the following conditions:

[0122] r i (t) = r i-1 (t)-IMF i (t) (3-4)

[0123] (h) Accumulate the IMFs obtained in step 3.1.1 (a)-(g), i.e., the loop conditions are i=i+1, i=1,2,…,k. This will give the final decomposition result.

[0124] Step 3.1.2: The residual component is defined as the remaining sequence in the final decomposition result, that is, the part that has not been decomposed.

[0125] Step 3.2: The ensemble empirical mode decomposition (EEMD) algorithm essentially applies EMD decomposition to the original load sequence for a given number of trials, then takes the average of all decomposition results. Unlike the EMD algorithm, Gaussian white noise with a uniform frequency distribution is added to the original sequence in each EEMD attempt. This ensures that the noise-assisted original load sequence not only has uniform distribution characteristics but also smoothness. Based on the initialization in step 3.1.1-(a), white noise of a given amplitude, σ, is added, specifically:

[0126] r0=x(t)+σ (3-5)

[0127]

[0128] Where A is the amplitude of the additive noise, σ is the white noise of a given amplitude, and K is the number of IMF components.

[0129] Repeat step 3.1 for a given number of experiments, with the number of IMFs remaining constant in each experiment. Finally, the current typical scenario is completed. The load time series is decomposed to obtain the eigenmodes used for prediction.

[0130] Step 3.3 is the same as step 3.1 and step 3.2, and the EEMD algorithm is applied to other typical scenarios. Perform empirical mode decomposition to obtain the IMF components and residual components under the corresponding scenario conditions.

[0131] Step 4: Decompose the historical load of different typical scenarios based on the EEMD algorithm applied in step 3, and reconstruct a new data set based on the weather, holiday and other characteristics in step 1.

[0132] Step 5: For the new data set of different heterogeneous load typical scenarios constructed in step 4, analyze the load characteristics under each scenario and build the corresponding load forecasting model. Adjust the hyperparameters of the neural network through grid search and optimize the objective function in combination with the Bayesian optimization algorithm, thus laying the foundation for accurate load forecasting under different scenarios. The principle of the Bayesian optimization algorithm is as follows Figure 6 shown.

[0133] Step 5.1: In-depth analysis of the characteristics of different load scenarios. Based on the load fluctuation patterns, periodicity, and trends of each scenario, identify the key factors that correspond to it.

[0134] Step 5.2: Design and construct corresponding load forecasting models (such as edRVFL, BiLSTM, SVR, etc.) based on the load characteristics of each scenario.

[0135] Step 5.3 combines the Bayesian optimization algorithm and uses the grid search method to adjust the hyperparameters of the neural network model and optimize the objective function of the load forecasting model. Bayesian optimization can efficiently search the hyperparameter space by constructing a proxy model and performing step-by-step optimization, thereby finding the global optimal solution under limited computing resources. The Bayesian optimization algorithm consists of five elements: proxy function, hyperparameter search space, acquisition function, objective function, and evaluation history. In the study, the objective function was defined as the prediction performance on the validation set. The popular Tree-based Parzen Window Estimation (TPE) algorithm was used to model the proxy function, and the expected improvement was used as the capture function, as shown in Formula (5-1):

[0136]

[0137] Among them, f is the objective function, f * is the threshold of the objective function for a given hyperparameter v.

[0138] Step 5.3.1 Initialization. Input parameter initialization: objective function f, TPE method Hyperparameter space Acquisition function verification Initialized memory Output parameter initialization: memory The best hyperparameter v within * .

[0139] Step 5.3.2 Iterative optimization. For each iteration h = 1 to H, use the TPE method Fitting Memory And calculate the probability distribution P(f|v) of the objective function under given hyperparameters.

[0140] Step 5.3.3: Maximize the acquisition function Find the next hyperparameter choice v i+1 , used to find a new hyperparameter combination from the hyperparameter space.

[0141] Step 5.3.4 Evaluate the objective function f(v i+1 ), calculate the new hyperparameter v i+1 The corresponding objective function value.

[0142] Step 5.3.5 Estimate the objective function f(v i+1 )Add storage Right now.

[0143] Step 5.3.6 Termination condition. Repeat the above steps 5.3.2 to 5.3.5 until all H iterations are completed, and finally output the best hyperparameter v found in memory D * .

[0144] Step 6: Complete the hyperparameter optimization iterations for the prediction model in Step 5 for different typical scenarios and build the corresponding model network. Input the new dataset constructed in Step 4, taking the first 80% as the training set and the first 20% as the test set. Fusion the prediction results for each component to ultimately derive a comprehensive forecast for each typical heterogeneous load scenario, and further, the overall load forecast for the park.

[0145] The above implementation scheme comprehensively considers the park's heterogeneous loads and multidimensional time-series variables, creatively incorporating them into the load forecasting model. Using the K-means clustering algorithm, the different types of heterogeneous loads within the park (such as industrial, commercial, and office buildings) are accurately classified, effectively identifying and modeling the characteristics of various load scenarios. This fine-grained analysis of heterogeneous loads ensures that the load forecasting model can adapt to the diversity of various types of loads within the park, improving the accuracy and reliability of forecasts.

[0146] Advantages of Adaptive Modeling and Intelligent Optimization: Utilizing the EEMD algorithm and Bayesian optimization techniques, we perform multi-scale decomposition of load sequences and intelligently optimize the hyperparameters of the forecast model, enabling adaptive adjustment of the forecast model to accommodate changes in different load scenarios. This dynamic adjustment process effectively enhances the model's flexibility and accuracy, enabling automatic optimization based on data characteristics, thereby improving load forecast accuracy and model robustness.

[0147] Compared with the existing technologies of "load forecasting input feature screening method based on hierarchical clustering and neural network" and "a load forecasting method considering multi-scale time series information", the present invention focuses on short-term load forecasting considering heterogeneous loads and multi-dimensional time series variables in the park, and can more accurately predict load demand in the context of power market reform and smart grid construction. First, missing values are filled in for historical load series, meteorological factors, holiday types, time of day, real-time electricity prices, and other factors, and outliers are removed. A K-means (Kmeans) clustering algorithm is then used to conduct an in-depth analysis of different types of heterogeneous loads within the park (such as industrial, commercial, and office buildings). By deeply mining the load data, accurate classification and modeling of various load scenarios in the park are achieved. Second, an ensemble empirical mode decomposition (EEMD) algorithm is used to decompose the park's heterogeneous load series, capturing the characteristics of signals in each frequency band. K intrinsic mode function (IMF) components and one residual component (RES) are obtained for loads of different scales. The input dataset for the prediction model is reconstructed by combining key information contained in multidimensional time series variables such as meteorological and holiday variables. Finally, targeted prediction models (edRVFL, BiLSTM, and SVR) are constructed for different load scenarios. The corresponding prediction model hyperparameters are optimized through grid search using a Bayesian optimization algorithm. The prediction results of each component are integrated to ultimately obtain a comprehensive prediction for each typical heterogeneous load scenario, and further, an overall load forecast for the park is obtained.

[0148] The following is a further demonstration and introduction of the present invention with reference to a specific test example:

[0149] The heterogeneous load forecasting process of the park based on EEMD and Kmeans adaptive optimization and the load forecasting model flow charts under different typical scenarios are shown in the following figure. Figure 1 and Figure 2 As shown, it mainly covers the following steps:

[0150] Step 1: Determine the park's heterogeneous load sequence, meteorological factors, holiday type, time and other factors collected by the system as sample characteristics, use the Lagrange interpolation method to fill in missing values, and select and eliminate outliers.

[0151] The heterogeneous load data for the park from 2022 to 2023, along with other influencing factors (including meteorological factors and holidays), was selected for research. All data was preprocessed at 30-minute intervals. Missing values were imputed through methods such as dropout, completion, and true value conversion, and outliers were selectively removed. Table 1 shows a subset of the preprocessed heterogeneous load data for the park (the last five entries).

[0152] Table 1 Preprocessed heterogeneous load data of some parks

[0153]

[0154] The input features of the preprocessed data (not yet time series decomposition) are shown in Table 2.

[0155] Table 2 Input characteristics of preprocessed data

[0156]

[0157] Step 2: Use the K-means clustering algorithm to perform cluster analysis on the pre-processed park heterogeneous load time series X, and divide it into multiple typical load scenarios based on the similarity of the load data to obtain the clustered load series The corresponding cluster labels are That is, they correspond to commercial load, industrial load, and optical cable load respectively.

[0158] Step 2.1 Randomly initialize the cluster center. For a given set of data Where n is the number of data samples, set to 35040; m is the dimension of the data, set to 10. Initialize the cluster centers to μ1, μ2, μ3, that is, let the number of clusters N = 3.

[0159] Step 2.2 calculates the Euclidean distance of each data point to the cluster center and assigns the data point to the cluster closest to it. i To each cluster center μ q The Euclidean distance is calculated and the cluster center with the closest distance is selected. The Euclidean distance formula is:

[0160]

[0161] Among them, xij For data x i The component in the jth dimension, μ qj is the cluster center μ q ,q=1,2,3 are the components in the jth dimension.

[0162] Then, for each data x i Classify it into the cluster closest to it q=1,2,3, that is:

[0163]

[0164] Step 2.3: Recalculate the cluster center based on the obtained cluster. The new cluster center is the mean of all data points in the cluster:

[0165]

[0166] in, It is a cluster The number of data points in .

[0167] Step 2.4 iteratively updates until the cluster center no longer changes or the change is less than the specified threshold. Repeat steps 2.2 and 2.3 until the change in the cluster center is less than the set threshold ε, that is:

[0168]

[0169] in, is the Euclidean distance of cluster center change, and the convergence threshold is set to ε = 0.001.

[0170] Step 3: Use EMD algorithm to analyze typical scenes The power load time series is stabilized to obtain the eigenmodes used for prediction.

[0171] Step 3.1 Apply EMD algorithm to The load time series is subjected to empirical mode decomposition to obtain five IMF components: i , i=1,2,3,4,5 and 1 residual component RES.

[0172] Step 3.1.1: Use the power load time series in step 2 as the initial input sequence and perform the following process on the initial sequence until the stopping criterion is met:

[0173] (a) Initialize r0=x1(t), i=1, h0=r i-1 (t), j = 1, where x(t) is the original sequence signal and h0 is the initial eigenmode component.

[0174] (b) By obtaining the sequence h j-1The local extreme points in the sequence are extracted by the maximum and minimum values of (t) to form the upper and lower envelopes.

[0175] (c) Calculate the average value of the upper and lower envelopes to obtain the mean function of the current sequence. The mean function formula is shown in formula (3-1):

[0176]

[0177] Among them, m j-1 (t) is the mean function of the j-1th iteration, up j-1 (t) is the upper envelope of the current iteration, low j-1 (t) is the lower envelope of the current iteration.

[0178] (d). From the original sequence h j-1 (t) Subtract the j-1th mean function to obtain the function h containing the high-frequency signal part j (t), h j (t) The formula is shown in formula (3-2):

[0179] h j (t) = h j-1 (t)-m j-1 (t) (3-2)

[0180] (e). Determine the current h j (t) Whether it is a qualified IMF. A qualified IMF must meet two conditions:

[0181] The number of extreme points and the number of zero crossings in the sequence are equal or differ by no more than 1.

[0182] Over the entire series, the average value of the amplitude at the extreme points is zero.

[0183] (f) If the currently obtained IMF does not meet the conditions of a qualified IMF, it is used as a new initial sequence, the loop condition j=j+1 is updated, and step 3.1.1-(b) is repeated.

[0184] (g) If the currently obtained IMF meets the conditions for a qualified IMF, it is saved, that is, the formula is satisfied:

[0185] IMF i (t) = h j (t) (3-3)

[0186] And subtract the IMF from the initial sequence, that is, satisfy the following conditions:

[0187] r i (t) = r i-1 (t)-IMF i(t) (3-4)

[0188] (h) Accumulate the IMFs obtained in steps (a)-(g) of step 3.1.1 to obtain the final decomposition result.

[0189] Step 3.1.2: The residual component is defined as the remaining sequence in the final decomposition result, that is, the part that has not been decomposed.

[0190] Step 3.2: Based on the initialization in step 3.1.1-(a), add white noise σ of a given amplitude, which is specifically expressed as:

[0191] r0=x(t)+σ (3-5)

[0192]

[0193] The amplitude of the additional noise is A=1, and the number of IMF components is K=5.

[0194] Repeat step 3.1 for a given number of experiments, with the number of IMFs remaining constant in each experiment. Finally, the current typical scenario is completed. The load time series is decomposed to obtain the eigenmodes used for prediction.

[0195] Step 3.3 is the same as step 3.1 and step 3.2, and the EEMD algorithm is applied to other typical scenarios. Performing empirical mode decomposition, we can also obtain 5 IMF components and 1 residual component under the corresponding scenario conditions.

[0196] Step 4: Decompose the historical load of different typical scenarios using the EEMD algorithm in step 3, and reconstruct a new dataset based on the weather, holiday, and other characteristics in step 1. The corresponding updated dataset (showing the last 5 entries of some data) is shown in Table 3.

[0197] Table 3 New heterogeneous load data of some parks after EEMD decomposition

[0198]

[0199] The input features corresponding to the updated data (time series decomposition has been performed, and the newly added features are shown in bold) are shown in Table 4.

[0200] Table 4 New data input features after EEMD decomposition

[0201]

[0202] Step 5: For the new dataset of typical heterogeneous load scenarios constructed in Step 4, analyze the load characteristics of each scenario and build a corresponding load forecasting model. Use grid search to adjust the neural network hyperparameters and combine the Bayesian optimization algorithm to optimize the objective function. The model hyperparameter adjustments for the corresponding scenarios are shown in Table 5.

[0203] Table 5 Hyperparameter optimization space for different models in typical scenarios

[0204]

[0205] Step 5.1: In-depth analysis of the characteristics of different load scenarios. Based on the load fluctuation patterns, periodicity, and trends of each scenario, identify the key factors that correspond to it.

[0206] Step 5.2: Design and construct corresponding load forecasting models (such as edRVFL, BiLSTM, SVR, etc.) based on the load characteristics of each scenario.

[0207] In step 5.3, the grid search method is used to adjust the hyperparameters of the neural network model in combination with the Bayesian optimization algorithm to optimize the objective function of the load forecasting model. In this study, the objective function uses RMSE as the prediction performance on the validation set, the popular tree-based Parzen window estimation (TPE) algorithm is used to model the surrogate function, and the expected improvement is used as the capture function, as shown in formula (5-1):

[0208]

[0209] Among them, f is the objective function, f * is the threshold of the objective function for a given hyperparameter v.

[0210] Step 5.3.1 Initialization. Input parameter initialization: objective function f, TPE method Hyperparameter space Acquisition function verification Initialized memory Output parameter initialization: memory The best hyperparameter v within * .

[0211] Step 5.3.2 Iterative optimization. For each iteration h = 1 to H, use the TPE method Fitting Memory And calculate the probability distribution P(f|v) of the objective function under given hyperparameters.

[0212] Step 5.3.3: Maximize the acquisition function Find the next hyperparameter choice v i+1, used to find a new hyperparameter combination from the hyperparameter space.

[0213] Step 5.3.4 Evaluate the objective function f(v i+1 ), calculate the new hyperparameter v i+1 The corresponding objective function value.

[0214] Step 5.3.5 Estimate the objective function f(v i+1 )Add storage Right now.

[0215] Step 5.3.6 Termination condition. Repeat the above steps 5.3.2 to 5.3.5 until all H iterations are completed and the memory is finally output. The best hyperparameter v found in * In this experiment, the optimal performance structure is determined by the optimization algorithm: for SVR, the kernel function type is determined to be rbf and the penalty parameter C = 2 -10 , the threshold ε of the loss function SVR =1×10 -3 Radius = 1 × 10 -2 and 32 hidden nodes; for the structural type hyperparameters of the Bi-LSTM neural network, determine the 4-layer network, Adam optimizer, 64 mini-batch size, 1×10 -3 The learning rate and the number of hidden nodes are 150, and the corresponding activation function is set to Tanh; for the structural type hyperparameters of the edRVFL neural network, the number of hidden nodes is 200, the network is 4 layers, the number of enhancement nodes is 20, the regularization parameter and input scaling are both 0, the ratio Ratios = 0.05, and the corresponding activation function is sigmoid.

[0216] Step 6: Complete the hyperparameter optimization iterations for the prediction model under different typical scenarios in Step 5 and build the corresponding model network. Input the new dataset constructed in Step 4, use the first 80% as the training set and the 20% as the test set, and fuse the prediction results of each component to ultimately obtain a comprehensive prediction for each typical heterogeneous load scenario, and further obtain the overall load forecast result for the park. The proposed model method is compared with the traditional single-model prediction SVR, Bi-LST, and edRVFL models (with the same optimized structure and training parameters). The corresponding model error evaluation indicators are shown in Table 6.

[0217] Table 6 Error evaluation indicators of the proposed prediction method compared with traditional SVR, Bi-LSTM and edRVFL

[0218]

[0219]

[0220] Through the analysis of the results, compared with the traditional prediction model (with the same optimization of structure and training parameters), the EEMD-Kmeans adaptive optimization load forecasting model that comprehensively considers the park's heterogeneous loads and multi-dimensional time series variables can more accurately predict load demand in the context of power market reform and smart grid construction.

[0221] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0222] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium having a computer program stored thereon, which executes the above method when executed by a processor. The storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component.

[0223] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0224] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present disclosure. Various changes and improvements may be made to the present disclosure without departing from the spirit and scope of the present disclosure, and such changes and improvements shall fall within the scope of the present disclosure.

[0225] The present invention is not limited to the above-mentioned optimal implementation mode. Anyone can derive various other forms of methods for determining rotor design parameters based on plateau environments under the guidance of the present invention. All equal changes and modifications made within the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A park heterogeneous load forecasting method based on EEMD and Kmeans adaptive optimization, characterized by: The following steps are involved: Fill missing values, remove outliers and perform feature coding on the park's historical load data and external influencing factors; Based on the similarity of load fluctuation characteristics, the park load is divided into multiple typical scenarios using the Kmeans clustering algorithm; Perform EEMD decomposition on the load sequence corresponding to each typical scenario, suppress modal aliasing by adding Gaussian white noise, extract multi-frequency components and fuse them with external variables to construct a prediction input data set; Select prediction models for different scenarios and adaptively tune them using the Bayesian optimization algorithm; The prediction results of each scenario are dynamically weighted and fused, and the weight is dynamically adjusted according to the historical load ratio of the scenario to output the overall load prediction value of the park.

2. The method for predicting heterogeneous load in a park based on EEMD and Kmeans adaptive optimization according to claim 1 is characterized by: The implementation of the Kmeans clustering specifically includes: The number of clusters is preset based on the park load type, and the initial cluster center is dynamically adjusted according to the historical load fluctuation curve; The scenario labels generated by clustering are directly used as the input of EEMD decomposition to ensure that the decomposition process only targets the same type of load fluctuation characteristics; The cluster centers are updated iteratively until the Euclidean distance variance of the load series within the cluster is less than the threshold.

3. The method for predicting heterogeneous load in a park based on EEMD and Kmeans adaptive optimization according to claim 1 is characterized by: The implementation of the EEMD decomposition specifically includes: The load sequence of each scenario is decomposed multiple times independently. Before each decomposition, Gaussian white noise of different instances is superimposed. The noise amplitude is inversely proportional to the square root of the number of decompositions. The decomposition process is dynamically bound to the classification results to ensure that the load fluctuation characteristics of industrial and commercial scenarios are independently characterized by differentiated frequency band components.

4. The method for predicting heterogeneous load in a park based on EEMD and Kmeans adaptive optimization according to claim 1, characterized in that: The Bayesian optimization adopts a tree-based Parzen window estimation strategy, the objective function is the root mean square error of the validation set, and the hyperparameters are selected by the expected improvement function.

5. The method for predicting heterogeneous load in a park based on EEMD and Kmeans adaptive optimization according to claim 1 is characterized by: The weight of the dynamic weight fusion is calculated based on the historical load ratio of each scenario.

6. The method for predicting heterogeneous load in a park based on EEMD and Kmeans adaptive optimization according to claim 1, characterized in that: The prediction models include edRVFL, BiLSTM, and SVR. The model type is selected based on the load fluctuation characteristics of the scenario: The edRVFL model is selected for industrial load scenarios; The BiLSTM model is selected for commercial load scenarios; Select the SVR model for optical cable load scenarios.

7. The method for predicting heterogeneous load in a park based on EEMD and Kmeans adaptive optimization according to claim 1, characterized in that: The external influencing factors include weather, holiday type, and real-time electricity price; The feature coding includes: Map weather conditions to numeric labels; The holiday type, day of the week, and time point are expanded to the Euclidean space through one-hot encoding.

8. A campus heterogeneous load forecasting system based on EEMD and Kmeans adaptive optimization, characterized by: include: Data preprocessing module: This module is configured to fill missing values, remove outliers, and perform feature encoding on the park's historical load data and external influencing factors, and output a standardized data set; Heterogeneous load classification module: This module is configured to use the Kmeans clustering algorithm to divide the campus load into multiple typical scenarios based on the similarity of load fluctuation characteristics, and output scenario labels and classified load sequences. Multi-scale signal decomposition module: This module is configured to perform EEMD decomposition on the load sequence of each scenario, suppress modal aliasing by adding Gaussian white noise, extract multi-frequency components and residual components, and fuse them with external variables to generate a multi-dimensional feature dataset; Scenario Adaptive Modeling and Optimization Module: This module is configured to select prediction models for different scenarios, adaptively tune model hyperparameters using the Bayesian optimization algorithm, and output the trained prediction model. Prediction result fusion module: It is configured to dynamically weight the prediction results of each scenario. The weight is calculated based on the historical load ratio of the scenario, and the overall load prediction value of the park is output.

9. The park heterogeneous load forecasting system based on EEMD and Kmeans adaptive optimization according to claim 8, characterized in that: The heterogeneous load classification module includes: Dynamic initialization unit: sets the initial number of clusters based on the preset park load type and dynamically adjusts the cluster center according to the historical load fluctuation curve; Iterative optimization unit: through Euclidean distance calculation and cluster center update, until the Euclidean distance variance of the load sequence within the cluster is less than the preset threshold; Classification-decomposition interface unit: Directly transmits the scene labels generated by clustering to the multi-scale signal decomposition module to ensure that the decomposition process is only for the same type of load data; The multi-scale signal decomposition module includes: Noise injection unit: superimposes Gaussian white noise on the classified load sequence, and the noise amplitude is inversely proportional to the square root of the decomposition number; Modal aliasing suppression unit: eliminates noise interference and extracts stable IMF components through multiple independent decompositions and result mean calculations; Feature binding unit: aligns the decomposed multi-frequency components with meteorological and holiday features by timestamp to construct a spatiotemporally correlated input dataset.

10. The park heterogeneous load forecasting system based on EEMD and Kmeans adaptive optimization according to claim 8, characterized in that: The scene adaptive modeling and optimization module includes: Model selection unit: Assigns prediction models based on the load fluctuation characteristics of the scenario. The edRVFL model is used for industrial load scenarios, the BiLSTM model is used for commercial load scenarios, and the SVR model is used for optical cable load scenarios. Bayesian optimization unit: uses the tree-based Parzen window estimation algorithm, takes the root mean square error of the validation set as the objective function, and searches for the optimal solution of hyperparameters through the expected improvement function; The prediction result fusion module includes: Weight calculation unit: dynamically adjusts weights based on the historical load ratio of the scenario; Fusion execution unit: performs weighted summation of the prediction values of each scene and iteratively optimizes the weight distribution through the error reverse correction mechanism.

Images