Plateau area load prediction method and system based on signal decomposition and neural network

Abstract

According to the plateau area load prediction method and system based on the signal decomposition and the neural network, an original signal is decomposed into a plurality of intrinsic mode components with physical significance by adopting completely adaptive noise set empirical mode decomposition, mode aliasing is suppressed, and multi-scale features are extracted to effectively suppress load demand oscillation; the problem of non-stationarity of the load of the service area is solved; then, K-means clustering grouping is carried out on the components, similar modes are combined to reduce calculation redundancy, and the problems of heterogeneity and local feature redundancy in complex time series data are effectively solved; performing nonlinear dimension reduction processing on the grouped reconstructed signals by adopting kernel principal component analysis, eliminating noise interference and extracting key features, thereby effectively solving the multiple nonlinear problem; and finally, inputting the dimensionality-reduced features into a long-short-term memory network modeling time sequence dependency relationship, and realizing high-precision prediction through adaptive weighted integration of component prediction results.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field related to load prediction, and particularly relates to a highland area load prediction method and system based on signal decomposition and neural networks. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] In recent years, with the deepening of the concept of "traffic + energy" integration development, the new type of service area energy system represented by the integration of "wind, light, storage and vehicle network" is being rapidly promoted, which puts forward higher requirements for load prediction technology. Highland areas have become strategic locations for clean energy development due to their abundant wind and solar energy resources. However, load prediction for highland highway service areas faces unique challenges. Influenced by extreme climate conditions and complex geographical environments, load data exhibits significant non-stationary, nonlinear and time-varying characteristics. Traditional prediction models are difficult to adapt to the special environment of highlands, resulting in large prediction errors, which in turn causes inaccurate scheduling of wind, light and storage systems, rising of curtailment rates and other problems, seriously restricting the efficient use of renewable energy. Conducting precise load prediction research for highland highway service areas not only helps to improve the green level of transportation infrastructure and promote the coordinated development of "traffic-energy" systems, but also provides technical support for high-proportion renewable energy grid connection.

[0004] Existing research mainly focuses on feature extraction of load sequence and optimization of prediction model, forming a relatively complete theoretical system. Some literature proposes a multi-energy load prediction method for highway service areas based on principal component modeling and improved fuzzy C-means clustering, which makes the prediction result time error less than 35%; some literature proposes a data and model driven state estimation method for highway service area traffic self-consistent energy system, which improves the prediction accuracy of photovoltaic output through a time convolution-bidirectional long short-term memory network model based on self-attention mechanism and predicts the spatial and temporal distribution of electric vehicle charging load based on Monte Carlo simulation; some literature proposes a highway service area electric vehicle charging load prediction method considering meteorological factors, which establishes an energy consumption model, analyzes the influence of weather on traffic flow, and combines geographic information systems and origin-destination matrix analysis method to predict the spatial and temporal distribution of charging load.

[0005] In the aspect of load forecasting model optimization, scholars have proposed various innovative methods. Some literature has constructed a hybrid neural network architecture combining convolutional neural networks and long short-term memory neural networks, achieving cross-modal fusion of meteorological multi-source heterogeneous data and load time series fluctuations. Some literature proposes a probabilistic density embedding type time series Transformer model, designs a conditional probability attention mechanism to model the historical load sequence as a Gaussian distribution. Some literature innovatively combines the whale optimization algorithm improved by Levy flight strategy with complete adaptive noise ensemble empirical mode decomposition signal decomposition, which makes the model convergence speed increase by 2.8 times, and achieves a prediction accuracy of 99.05% on the high-altitude service area dataset. Some literature proposes a two-stage adaptive signal decomposition coupled with spatio-temporal attention mechanism prediction framework, which significantly improves the prediction performance during typhoon passage. In addition, some literature proposes a multi-task learning method based on multi-dimensional fusion features and convolutional neural networks, which effectively improves the operation efficiency and accuracy of massive user short-term load forecasting. Some literature constructs a framework that integrates isolated forest anomaly correction, variational mode decomposition component extraction, and multilayer perceptron-gated recurrent unit sub-item prediction, enhancing spatial load situation awareness capability. Some literature combines twin network similar day classification and grey wolf algorithm to optimize long short-term memory neural network parameters, improving the accuracy and efficiency of short-term load forecasting of distribution networks. Some literature uses bidirectional long short-term memory neural network to predict load key values, constructs a similar curve dataset through weighted K- nearest neighbor method, and generates complete load curve using conditional generative adversarial network, combining with probability mapping to achieve high-precision short-term probability prediction. In addition, some literature proposes time-varying filter empirical mode decomposition, which reorganizes intrinsic mode function (IMF) components by calculating sample entropy, reducing prediction error to below 3%. Some literature uses variational mode decomposition to separate low-frequency trends and high-frequency fluctuations in load sequences, and combines them with meteorological factors to input into a hybrid deep learning model combining convolutional neural networks and bidirectional gated recurrent unit hybrid model.

[0006] However, the above studies are mainly aimed at plain area scenes, and there are obvious limitations when applied to plateau service areas. Plateau areas have the characteristics of high proportion of new energy, complex climate and environment, etc., resulting in more prominent load volatility and uncertainty than plain areas, and traditional prediction methods are difficult to effectively cope with this challenge. The coupling of extreme climate mutations and complex road network topology in the plateau leads to non-stationarity and multi-modal characteristics of charging sequences far exceeding those in the plain area. Traditional modal decomposition still has difficulty in completely suppressing the modal aliasing problem caused by the end effect when dealing with such dramatic fluctuations. The geographical barriers in the plateau result in sparse distribution of charging stations, and user behavior data presents high-dimensional sparsity. Conventional linear dimensionality reduction methods or feature extraction methods are difficult to effectively analyze the complex nonlinear implicit relationship between key factors such as altitude and charging power. More importantly, existing neural network prediction models generally lack deep coupling with plateau-specific physical mechanisms, making it difficult for the prediction model to accurately capture the dynamic evolution law of plateau charging load, and the precision is difficult to meet the fine regulation and control demand of plateau power grid.

[0007] In summary, based on the differences in load characteristics between plain areas and plateau areas, how to accurately predict the load in plateau areas based on the characteristics of the load in plateau areas is a problem that needs to be solved. SUMMARY

[0008] To overcome the shortcomings of the prior art, the present application provides a plateau area load prediction method and system based on signal decomposition and neural networks, and proposes a CEEMDAN-K-means clustering-KPCA-LSTM hybrid prediction model to achieve accurate prediction of load changes in plateau areas.

[0009] To achieve the above purpose, the technical scheme adopted by the present application is as follows: In a first aspect, the present application provides a plateau area load prediction method based on signal decomposition and neural networks, comprising: The original load is decomposed into a plurality of intrinsic mode components by complete adaptive noise ensemble empirical mode decomposition; The plurality of intrinsic mode components obtained are subjected to clustering analysis by a K-means clustering algorithm, and similar intrinsic mode components are combined to obtain reconstructed intrinsic mode components; The reconstructed intrinsic mode components are subjected to nonlinear dimensionality reduction processing by kernel principal component analysis to extract key features; The extracted key features are input into a trained long short-term memory network to model the time series dependency relationship, and the load prediction result is obtained.

[0010] In a second aspect, the present application provides a plateau area load prediction system based on signal decomposition and neural networks, comprising: The decomposition module is configured to decompose the original load into a plurality of intrinsic mode components by complete adaptive noise ensemble empirical mode decomposition; The clustering module is configured to perform clustering analysis on the obtained multiple intrinsic modal components by using a K-means clustering algorithm, and to merge similar intrinsic modal components to obtain reconstructed intrinsic modal components. The dimension reduction module is configured to perform nonlinear dimension reduction processing on the reconstructed intrinsic modal components by using kernel principal component analysis, and to extract key features. The prediction module is configured to input the extracted key features into a trained long short-term memory network to model the time series dependency relationship and obtain a load prediction result.

[0011] In a third aspect, the present application provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein when the computer instructions are executed by the processor, the method of the first aspect is completed.

[0012] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method of the first aspect is completed.

[0013] The above one or more technical solutions have the following beneficial effects: In the present application, the original signal is decomposed into multiple intrinsic modal components with physical meaning by using completely adaptive noise ensemble empirical mode decomposition, the modal aliasing is suppressed, and the multi-scale features are extracted to effectively suppress the load demand shock, and the non-stationary problem of the service area load is solved; then the components are grouped by K-means clustering, and similar modes are merged to reduce the calculation redundancy, effectively solving the heterogeneity and local feature redundancy problem in complex time series data; then the grouped reconstructed signal is processed by nonlinear dimension reduction by kernel principal component analysis, the noise interference is eliminated, and the key features are extracted, effectively solving the multiple nonlinear problems; finally, the features after dimension reduction are input into the long short-term memory network to model the time series dependency relationship, and high-precision prediction is realized through adaptive weighted integration of the component prediction results.

[0014] In the present application, in order to optimize the feature expression of the high-dimensional IMF component after decomposition, a component reorganization strategy based on sample entropy and K-means clustering is proposed, the sample entropy of each IMF is calculated to quantify the complexity feature, and then the K-means algorithm is used to divide it into three categories of high-frequency noise, medium-frequency periodic component and low-frequency trend component, so as to enhance the feature representativeness and improve the quality of LSTM modeling.

[0015] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification. The embodiments of the application, and their

[0017] Figure 1 LSTM principle structure diagram in embodiment one of the application; Figure 2 Flow chart of highland area load prediction method based on signal decomposition and neural network in embodiment one of the application; Figure 3 Schematic diagram of CEEMDAN decomposition result in embodiment one of the application; Figure 4 Schematic diagram of IMFs reconstruction result in embodiment one of the application; Figure 5 Comparison chart of three model predictions in embodiment one of the application. DETAILED DESCRIPTION

[0018] It should be noted that the following detailed description is merely exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0019] It should be noted that the terms used herein are merely for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application.

[0020] In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0021] Embodiment one The embodiment discloses a highland area load prediction method based on signal decomposition and neural network, comprising: The original load is decomposed into a plurality of intrinsic mode components by complete self-adaptive noise ensemble empirical mode decomposition; The plurality of intrinsic mode components obtained are subjected to clustering analysis by using K-means clustering algorithm, and similar intrinsic mode components are combined to obtain reconstructed intrinsic mode components; The reconstructed intrinsic mode components are subjected to nonlinear dimensionality reduction processing by using kernel principal component analysis to extract key features; The extracted key features are input into the trained long short-term memory network to model the time series dependency relationship, and load prediction results are obtained.

[0022] The embodiment constructs a hybrid prediction architecture combining signal decomposition and neural network, and innovatively integrates the K-means method into the data preprocessing part, effectively solving the problems of heterogeneity and local feature redundancy in complex time series data. The embodiment first briefly describes the influence of plateau climate on service area load prediction, which causes the plateau service area load to have the characteristics of strong non-stationarity, nonlinearity and time heterogeneity. In view of the unique characteristics of service area load, the original signal is decomposed into multiple intrinsic mode components with physical meaning by using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), which suppresses mode mixing and extracts multi-scale features to effectively suppress load demand shock. Then, based on the statistical property of sample entropy, the components are K-means clustered and grouped, and similar modes are merged to reduce computational redundancy; then the reconstructed signal after grouping is processed by kernel principal component analysis (KPCA) for nonlinear dimensionality reduction, eliminating noise interference and extracting key features; finally, the reduced features are input into the long short-term memory network (LSTM) to model the time series dependence relationship, and the adaptive weighted integration of component prediction results is realized to achieve high-precision prediction. The measured data of Riyueshan service area in Qinghai Province are used for verification, which verifies the feasibility and accuracy of the model.

[0023] The unique climate conditions in plateau areas have a significant impact on the characteristics of service area load. The characteristics of low oxygen, low pressure, strong solar radiation, and large diurnal temperature difference directly change the operation mode and energy consumption law of service area infrastructure.

[0024] The plateau highway service area power load presents typical non-stationary characteristics, which is mainly due to the influence of special geographical and climatic environment. In the plateau area, the diurnal temperature difference caused by the thin atmosphere is very large, which makes the heating and cooling system need to frequently switch the operating state, and the daily load curve presents obvious sawtooth fluctuation. At the same time, lightning activities are frequent in the plateau area in summer, and voltage sag caused by lightning may often lead to automatic switching of standby power supply, which will form a transient peak on the load curve. The low-pressure hypoxic environment reduces the heat dissipation efficiency of electrical equipment, which not only raises the basic load level, but also causes the response time of temperature control equipment to be prolonged, further amplifying the load fluctuation amplitude.

[0025] The service area load change presents a significant nonlinear characteristic, which is mainly reflected in the complex coupling relationship of multiple influencing factors. The measured data show that when the solar radiation intensity exceeds When the temperature difference between indoor and outdoor is large, the heat transfer through the enclosure of the service area will cause the nonlinear growth of air conditioning load. The activity of people also shows a special nonlinearity. At high altitudes, the average stay time of passengers will be shortened due to the decrease of blood oxygen saturation, which will make the peak time of catering and sanitation facilities advance. More complex is the impact of traffic control. When traffic control is implemented due to severe weather, stranded vehicles will quickly gather in a short time, causing the load of the service area to rise sharply in a short time, and the load will decrease after the traffic control is lifted, showing obvious hysteresis characteristics. This asymmetric response process brings great challenges to load forecasting.

[0026] The load of the plateau service area shows significant temporal heterogeneity in time distribution, which is characterized by the dynamic superposition and complex coupling of multi-time scale load characteristics. On the scale of seconds, the unique meteorological conditions in the plateau region are prone to lightning disasters. The transient overvoltage and overcurrent caused by lightning strikes will cause millisecond to second level pulse load in the power supply system of the service area, which has a magnitude far exceeding the normal power consumption level, and has randomness and suddenness, which poses a threat to the safe operation of power equipment. On the scale of hours, there are obvious differences in passenger travel activities. In the morning and at night, passenger flow is small, and the power demand of charging piles and lighting equipment is low. During the passenger flow peak in the afternoon and evening, the power load of catering equipment, hot water supply, and charging service rises sharply, showing obvious periodic fluctuations. In terms of seasonal scale, the plateau climate conditions are complex and changeable. During the summer tourist season, a large number of tourists pour in, causing various facilities in the service area to run at full capacity and the power demand to surge. In winter, the heating equipment runs for a long time, and the charging efficiency is affected by the low temperature, which increases the charging time, making the overall power consumption structure and load level completely different from that in summer. These different time scale load characteristics are intertwined, further increasing the complexity and heterogeneity of the plateau service area load, and posing a severe challenge to traditional prediction methods.

[0027] In view of the strong non-stationarity and nonlinearity of the plateau service area load, the CEEMDAN method is used to decompose the original data in this embodiment. Through the iterative processing of adaptive noise injection and ensemble average, the load characteristic components of different time scales are effectively separated. CEEMDAN is a new signal decomposition method proposed to solve the problems of mode mixing, noise residue, and low computational efficiency of traditional empirical mode decomposition. The core idea of CEEMDAN is to realize accurate decomposition of strong non-stationary and nonlinear signals through adaptive noise injection and complete set construction. CEEMDAN algorithm inherits the adaptive decomposition characteristics of empirical mode decomposition, and effectively suppresses the mode mixing phenomenon by introducing adaptive Gaussian white noise and multiple set average strategy.

[0028] Under the plateau climate effect, the time-frequency characteristics of load sequence present special rules. The load change affected by the high-frequency characteristics of intraday photovoltaic output fluctuation caused by strong radiation is mainly reflected in the high-frequency intrinsic mode function component, while the load change affected by the low-frequency characteristics of equipment maintenance load caused by seasonal freeze-thaw cycle is concentrated in the low-frequency intrinsic mode function. Extreme climate events can cause modal aliasing phenomenon, and specific climate impact exists in adjacent intrinsic mode function components at the same time.

[0029] The basic steps are as follows: Add Gaussian white noise to the original signal to generate a perturbed signal: (1) In the formula, is the perturbed signal generated after the i-th Gaussian white noise is added, is the original signal, N is the standard deviation of noise, which is used to control the intensity of added noise, is the i-th noise sequence, i represents the number of added noise, i=1, 2, …, NR, and NR is the total number of added noise.

[0030] Modal decomposition is performed on each noisy signal to obtain the average value of the first-order IMF and update the residual error: (2) (3) In the formula, is the average value of the first-order intrinsic mode function obtained by performing empirical mode decomposition on NR noisy signals. The first IMF obtained by modal decomposition of the i-th noisy signal is The residual error signal obtained by subtracting the average value of the first-order IMF from the original signal is Extract the k-th order IMF ( , ) Repeat the following process for the current residual error : Add noise to the residual error signal: (4) In the formula, is the signal after the i-th noise is added based on the residual error signal of the k-1 step, is the residual error signal obtained by the k-1 step decomposition, is the energy normalization coefficient of the first k-1 order IMF.​​

[0031] Perform mode decomposition on the noisy residuals, take the average value of the first IMF, and update the residuals: (5) (6) In the formula, The k-th order eigenmode function is obtained by performing a series of operations on the residual signal. To the residual signal after adding noise The first IMF obtained by performing EMD decomposition, For the k-th step decomposition, from Subtract The residual signal obtained afterwards.

[0032] When the residual Decomposition stops when the function is monotonic or when the energy is below a threshold. The final residual is the trend term. .

[0033] Since the CEEMDAN decomposition yields too many IMF components, resulting in excessive computational burden, this embodiment employs K-means clustering to perform feature clustering on the IMFs obtained from CEEMDAN decomposition, merging IMFs with similar features to reduce data complexity. By calculating the sample entropy of each IMF component as a clustering feature, and using the silhouette coefficient to determine the optimal number of clusters, the complex nonlinearly coupled modes are divided into subsets of modes with similar characteristics.

[0034] Traditional methods that directly use all IMF components can lead to feature redundancy and noise interference, while K-means clustering can achieve modality grouping, enhance feature representativeness, and improve the quality of LSTM modeling.

[0035] For data processed by K-means clustering, this example uses the KPCA method to extract core features. KPCA uses kernel tricks to nonlinearly map the original feature space to a high-dimensional Hilbert space, thereby achieving linear separability transformation of complex modal features. Traditional linear dimensionality reduction methods struggle to capture the nonlinear relationship between abrupt changes in temperature and humidity and load response, while KPCA, using a Gaussian kernel function, can effectively characterize the complex interaction between climate parameters and load features.

[0036] Its mathematical essence can be expressed as: given the intrinsic mode function matrix after K-means clustering and CEEMDAN reconstruction L represents the number of monitored samples, N represents the number of IMF components, and the constructed plateau adaptive Gaussian kernel matrix is... The element is defined as: (7) In the formula, The IMF feature vector representing the i-th sample; The IMF feature vector representing the j-th sample; Represents the Kronecker symbol, which is 1 when i=j and 0 otherwise; Represents altitude-adaptive kernel width; This represents the regularization coefficient.

[0037] The kernel parameter σ employs an altitude-adaptive strategy. (8) In the formula, Represents the baseline core width; H This represents the altitude of the monitoring point.

[0038] For the centralized kernel matrix: (9) In the formula, It represents a column vector consisting entirely of 1s.

[0039] Perform eigenvalue decomposition Then, select the previous one. m The largest eigenvalues The corresponding feature vectors ultimately yield a low-dimensional feature projection with adaptability to the plateau environment: (10) In the formula, V Λ represents the eigenvector matrix; Λ represents the diagonal eigenvalue matrix. This represents the projection of low-dimensional features.

[0040] To address the temporal heterogeneity of load in plateau service areas, this embodiment employs a Long Short-Term Memory (LSTM) neural network to process data processed by CEEMDAN, K-means clustering, and KPCA. As an improved architecture of recurrent neural networks, LSTM effectively solves the gradient vanishing and exploding problems inherent in traditional recurrent neural networks in temporal modeling through gating mechanisms and cell state transfer strategies. It is particularly suitable for prediction tasks with strong temporal dependencies and nonlinear dynamic characteristics.

[0041] The core mechanism of Long Short-Term Memory (LSTM) neural networks lies in constructing a cell state update path controlled collaboratively by a forget gate, an input gate, and an output gate. The forget gate utilizes the sigmoid function. Calculate historical states With the current input weight vector Determines cell state The degree of information retention in the input gate; the input gate then combines the candidate states. With update coefficient , screen the effective components in the current time series features. Both of them update the cell state , realize the dynamic adjustment of long-term memory; the output gate generates the hidden state based on the updated , extract the time series pattern with physical meaning through tanh nonlinear transformation. The gating mechanism can effectively capture the time evolution law of highland climate. The input gate filters historical climate information through the sigmoid function, the forget gate dynamically adjusts the long-term dependence of device thermal inertia, and the output gate generates the load prediction value by synthesizing the current climate state. In the scenario of extreme weather on the plateau, the long short-term memory neural network stores the persistent influence of climate events through the cell state.

[0042] In view of the non-stationarity, multiple nonlinearities and time heterogeneity of plateau service area load, a CEEMDAN-K-means clustering-KPCA-LSTM coupled model is proposed. This model combines multi-stage signal decomposition, clustering simplification, feature dimension reduction and deep time series modeling to achieve high-precision prediction of non-stationary time series data. The steps are as follows: 1. The data preprocessing stage adopts a sliding window dynamic detection algorithm to define a time-varying detection window based on local statistics to construct an adaptive threshold function and correct outliers. In view of the spatial heterogeneity of plateau highway service areas, for the case where the average distance between two service areas is km, the net power is used as the input feature for the section between the two service areas with a distance greater than 15 km.

[0043] 2. The CEEMDAN decomposition method is introduced to suppress the modal aliasing effect through an adaptive noise injection mechanism, and each column of signal is decomposed into multiple IMF and 1 residual component. Each IMF component has quasi-stationary characteristics, thus solving the non-stationarity problem of service area load.

[0044] 3. To optimize the feature expression of high-dimensional IMF components after decomposition, a component reorganization strategy based on sample entropy and K-means clustering is proposed. First, calculate the sample entropy of each IMF to quantify the complexity feature, and then use the K-means algorithm to divide it into three categories: high-frequency noise, medium-frequency periodic component and low-frequency trend component.

[0045] 4. Further, KPCA method is used to reduce the nonlinear features of the reorganized components. Through the Gaussian kernel function mapping to high-dimensional space, the principal components with cumulative contribution rate above 95% are extracted, effectively solving the multiple nonlinear problems.

[0046] 5. In the data standardization stage, the input features are normalized by maximum and minimum values, the output load values ​​are standardized by Z-score to eliminate dimensional differences, and the temporal dependence is enhanced by time window reconstruction.

[0047] 6. By training the LSTM model, temporal feature fusion and climate response modeling are achieved, thereby improving the accuracy of model predictions.

[0048] The following examples demonstrate the effectiveness of the plateau load forecasting method based on signal decomposition and neural networks proposed in this embodiment: To comprehensively evaluate the predictive performance of the CEEMDAN-K-means clustering-KPCA-LSTM model, multi-dimensional quantitative indicators were selected for comprehensive analysis. These included root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination. ).

[0049] Root mean square error (RMSE): It reflects the overall distribution characteristics of prediction error by calculating the square root of the square mean of the deviations between the predicted and actual values, and has a high sensitivity to outliers.

[0050] (11) in, These are actual observed values. These are the model's predicted values. The total number of samples.

[0051] Mean Absolute Error: The mean absolute error characterizes the stability of the prediction results, intuitively reflects the average fitting accuracy of the model to time series fluctuations, and reflects the scale of the absolute error.

[0052] (12) Mean absolute percentage error: This measure of relative error in percentage form effectively eliminates the influence of differences in data dimensions on the evaluation results. It is suitable for cross-sectional comparative analysis of multi-scale sequences and characterizes the distribution characteristics of relative error.

[0053] (13) The coefficient of determination (COD) quantifies the model's ability to explain the variance of the target variable from a statistical perspective. The closer the COD value is to 1, the stronger the model's ability to capture the inherent patterns of time series. The mathematical expression for it is defined as: (14) This indicator reflects the model's explanatory power for data variance by the ratio of the residual sum of squares to the total sum of squares, and its value range is [value range missing]. .when This indicates that the predicted value perfectly matches the actual value. The positive value of the model prediction ability is equivalent to the direct use of the mean value prediction, and the negative value means that the model fitting effect is lower than the benchmark mean line. This index is often used to quantify the reconstruction accuracy of the time series fluctuation characteristics.

[0054] The embodiment takes the load data of the Riyueshan service area power grid in Qinghai Province from July 2024 to December 2024 as the research object, selects historical load, temperature, traffic flow, altitude, solar radiation, humidity, oxygen content, wind speed, and ultraviolet intensity as the original data, and constructs a time series data set based on a 15-minute sampling interval. In view of the non-stationary, nonlinear, and time-heterogeneous characteristics of the load signal of the plateau service area, a combined data feature processing method is proposed, which fuses CEEMDAN, K-means clustering, and KPCA.

[0055] The CEEMDAN decomposition of the training sample can obtain a series of IMF components, such as Figure 3 The CEEMDAN decomposition result is shown in FIG. 4.

[0056] It can be seen that each IMF component presents obvious frequency distribution characteristics from high to low. The IMF1-IMF5 components have intense fluctuations and short periods, mainly reflecting the high-frequency noise and random fluctuation components in the load signal. The IMF6-IMF9 components have increased amplitudes and more obvious periodic characteristics, corresponding to the medium-frequency characteristics of the load change, which may be related to the daily periodic fluctuation of the service area traffic flow. The IMF10-IMF13 components exhibit slow-changing trend characteristics, reflecting the low-frequency rules of the load affected by meteorological factors and other long-term influences. The residual term presents a monotonic increasing trend, implying that the service area load may have a long-term growth trend. This multi-scale decomposition result effectively reveals the non-stationary characteristics of the plateau service area load signal, providing an important basis for subsequent feature reconstruction and prediction modeling. There is no obvious modal aliasing between the IMF components, indicating that the CEEMDAN method can effectively overcome the limitations of the traditional EMD algorithm under the complex climate conditions of the plateau, and realize the clear separation of the load signal in different time scales.

[0057] After calculating the sample entropy of each IMF, the IMF is reconstructed using K-means clustering to obtain high-frequency, medium-frequency, and low-frequency components as shown in FIG. 4.

[0058] It can be seen that after K-means, the IMF will be divided into high-frequency components, medium-frequency components, and low-frequency components. Among them, the high-frequency components correspond to IMF1-IMF5 after CEEMDAN decomposition; the medium-frequency components correspond to IMF6-IMF9; and the low-frequency components correspond to IMF10-IMF13.

[0059] The high-frequency, medium-frequency and low-frequency components reconstructed by K-means clustering are further subjected to kernel principal component analysis for feature extraction. This hierarchical feature extraction method not only significantly reduces the data dimension, but more importantly, effectively reveals the complex nonlinear relationship between the plateau service area load and the multi-source influencing factors through kernel function mapping, providing more physically meaningful feature input for the subsequent construction of the prediction model. Experiments show that the feature set processed by KPCA retains more than 95% of the information of the original data, while improving the model training efficiency by about 30%.

[0060] In order to further verify the effectiveness of the CEEMDAN-K-means clustering-KPCA-LSTM model, the embodiments compare and verify the LSTM method, the CEEMDAN-LSTM method and the CEEMDAN-K-means clustering-KPCA-LSTM method, and the comparison chart is shown in FIG. 5.

[0061] Through error analysis, the evaluation indexes of the three models can be obtained, as shown in Table 1.

[0062] Table 1: Comparison of evaluation indexes of three models

[0063] From Figure 4 From the comparison chart and the comparison of the evaluation indexes in Table 1, it can be concluded that the proposed CEEMDAN-K-means clustering-KPCA-LSTM has improved evaluation indexes compared with the other two methods, and it has strong prediction ability for the load under plateau conditions The embodiments innovatively construct a joint prediction model that combines CEEMDAN, K-means, KPCA and long short-term memory network LSTM, aiming at the non-stationary, nonlinear and time-heterogeneous characteristics of the plateau service area load sequence. The CEEMDAN-K-means clustering-KPCA-LSTM hybrid prediction model is proposed, which realizes the accurate prediction of load changes.

[0064] In the embodiments, CEEMDAN is used to decouple the modal of multi-dimensional meteorological-load data on the plateau, and the adaptive noise injection strategy is used to solve the modal confusion problem existing in traditional empirical mode decomposition. Then, based on the statistical characteristics of sample entropy, the components are subjected to K-means clustering grouping, which effectively reduces the calculation redundancy. Combined with KPCA, a nonlinear feature space mapping is constructed, and the principal components with a cumulative contribution rate of 95% are extracted from the original features, effectively solving the "dimension disaster" problem in the sparse data environment on the plateau.

[0065] The embodiment utilizes the LSTM method for modeling simulation, and the actual measurement data verification shows that the RMSE, MAE and MAPE of the CEEMDAN-KPCA-LSTM model reach 100.611 kW, 85.9372 kW and 1.9451%, which is about 17.8% higher than the accuracy of the single LSTM model; compared with the CEEMDAN-LSTM combined model, it is still improved by 12.7%.

[0066] The embodiment reduces the standard deviation of the model prediction error by introducing the ultraviolet intensity specific to the plateau, the freeze-thaw cycle days and the vehicle flow and other multi-source characteristic parameters. Research proves that the mixed framework effectively solves the dimension disaster problem in the sparse data environment of the plateau, and provides a generalizable technical path for the accurate scheduling of energy systems in high-altitude areas.

[0067] Embodiment two The purpose of the embodiment is to provide a plateau area load forecasting system based on signal decomposition and neural network, comprising: The decomposition module is configured to decompose the original load into a plurality of intrinsic mode components by using the completely self-adaptive noise ensemble empirical mode decomposition; The clustering module is configured to perform clustering analysis on the plurality of intrinsic mode components obtained by using the K-means clustering algorithm, and combine similar intrinsic mode components to obtain reconstructed intrinsic mode components; The dimension reduction module is configured to perform nonlinear dimension reduction processing on the reconstructed intrinsic mode components by using kernel principal component analysis to extract key features; The prediction module is configured to input the extracted key features into the trained long short-term memory network to model the time series dependence relationship and obtain the load prediction result.

[0068] In more embodiments, there are also provided: An electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the method described in embodiment one is completed. For brevity, it will not be repeated here.

[0069] It should be understood that in the embodiment, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, ready-to-program gate arrays FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0070] The memory can include read-only memory and random access memory, and provide instructions and data to the processor, a portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0071] A computer readable storage medium for storing computer instructions, which are executed by a processor to complete the method described in embodiment one.

[0072] The method in embodiment one can be directly embodied as a hardware processor to complete, or be completed by a combination of hardware and software modules in the processor. The software modules can be located in the storage medium mature in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, register, etc. The storage medium is located in the memory, and the processor reads the information in the memory to complete the steps of the above method in combination with the hardware. To avoid repetition, it will not be described in detail here.

[0073] A computer program product comprising a computer program, which, when executed by a processor, implements the method described in embodiment one.

[0074] The present application also provides at least one computer program product tangibly stored on a non-transitory computer readable storage medium. The computer program product includes computer executable instructions, such as those included in program modules, executed by devices at a destination real or virtual processor to perform processes / methods as described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In various embodiments, the functionality of program modules can be combined or split between program modules as desired. Machine executable instructions for program modules can be executed within a local or distributed device. In a distributed device, program modules can be located in local and remote storage media.

[0075] Computer program code for carrying out operations of the present application can be written in one or more programming languages. These computer program codes can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, so that the program codes cause the functions / operations specified in the flowcharts and / or block diagrams to be performed when the computer or other programmable data processing apparatus executes the program codes. The program codes can be executed entirely on the computer, partially on the computer, as a standalone software package, partially on the computer and partially on a remote computer, or entirely on a remote computer or server.

[0076] In the context of the present application, the computer program code or related data can be carried by any suitable carrier to enable the device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals can include electrical, optical, radio, sound or other forms of propagated signals, such as carrier waves, infrared signals, and the like.

[0077] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the present embodiment can be realized in electronic hardware or in combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0078] The above describes the specific embodiments of the present application in conjunction with the drawings, but is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A load forecasting method for highland areas based on signal decomposition and neural networks, characterized by, The method comprises the steps of: using fully adaptive noise ensemble empirical mode decomposition to decompose the original load into multiple intrinsic mode components; using a K-means clustering algorithm to perform clustering analysis on the multiple intrinsic mode components obtained, merging similar intrinsic mode components to obtain reconstructed intrinsic mode components; using kernel principal component analysis to perform nonlinear dimensionality reduction processing on the reconstructed intrinsic mode components, and extracting key features; inputting the extracted key features into a trained long short-term memory network to model the time series dependency relationship and obtain a load prediction result. 2.The highland area load forecasting method based on signal decomposition and neural network according to claim 1, wherein, The method comprises the steps of: adding Gaussian white noise to the original load signal to generate a perturbed load signal; performing mode decomposition on each noise-added signal to obtain the average value of the first-order intrinsic mode component and update the residual error; extracting the kth-order intrinsic mode component, adding noise to the current residual error, performing mode decomposition on the noise-added residual error, taking the average value of the first intrinsic mode component and updating the residual error, and stopping the decomposition when the residual error is a monotonic function or the energy is lower than a threshold value. 3.The highland area load forecasting method based on signal decomposition and neural network of claim 1, wherein, The method comprises the steps of: calculating the sample entropy of each intrinsic mode component as a clustering feature; using the silhouette coefficient to determine the optimal number of clusters, dividing the complex mode with nonlinear coupling into mode subsets with similar characteristics, merging similar intrinsic mode components, and obtaining reconstructed intrinsic mode components. 4.The highland area load forecasting method based on signal decomposition and neural network according to claim 1, wherein, The method comprises the steps of: constructing a Gaussian kernel matrix for the reconstructed intrinsic mode component matrix, performing eigenvalue decomposition on the centralized kernel matrix, selecting the orthogonal feature vectors corresponding to the first m largest eigenvalues, and obtaining low-dimensional principal component projection. 5.The highland area load forecasting method based on signal decomposition and neural network according to claim 1, wherein, Before using fully adaptive noise ensemble empirical mode decomposition to decompose the original load into multiple intrinsic mode components, the method further comprises the steps of: 6.The highland area load forecasting method based on signal decomposition and neural network according to claim 4, wherein, using a sliding window method combined with a 3σ criterion to dynamically detect outliers, performing linear interpolation correction on outliers that exceed the set range, and ensuring time series continuity through missing value filling.

7. A highland area load forecasting system based on signal decomposition and neural networks, characterized by, According to the altitude of the monitoring point, an altitude-adaptive kernel parameter function is constructed; and the Gaussian kernel matrix is constructed based on the altitude-adaptive kernel parameter function. The method comprises the steps of: a decomposition module configured to use fully adaptive noise ensemble empirical mode decomposition to decompose the original load into multiple intrinsic mode components; a clustering module configured to use a K-means clustering algorithm to perform clustering analysis on the multiple intrinsic mode components obtained, merge similar intrinsic mode components, and obtain reconstructed intrinsic mode components; a dimensionality reduction module configured to use kernel principal component analysis to perform nonlinear dimensionality reduction processing on the reconstructed intrinsic mode components and extract key features; a prediction module configured to input the extracted key features into a trained long short-term memory network to model the time series dependency relationship and obtain a load prediction result.

8. An electronic device, comprising: A computer program product comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, which, when run by the processor, complete the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, A computer program product for storing computer instructions, which, when executed by a processor, complete the method of any one of claims 1-6.

10. A computer program product, characterised in that, A computer program product comprising a computer program, which, when executed by a processor, implements the method of any one of claims 1-6.

Images