Distributed new energy load prediction method based on complete set empirical mode decomposition

Abstract

The invention discloses a distributed new energy load prediction method based on complete set empirical mode decomposition, and relates to the technical field of new energy load prediction.The method comprises the steps that basic data are preprocessed, the basic data are decomposed into a plurality of intrinsic mode functions and remainders through complete set empirical mode decomposition, and based on the decomposed intrinsic mode functions and remainders, a new energy load is predicted; constructing an optimized feature set according to feature importance output by the extreme gradient boosting tree; based on the optimized feature set, training the long-term dependency type subsequences by using a long short-term memory network, training the short-term fluctuation type subsequences by using an extreme learning machine, and carrying out scale-division modeling to obtain load prediction values of the subsequences; and superposing the load prediction values of the subsequences to obtain a final power prediction result. The method is excellent in prediction precision, can adaptively decompose a power sequence and screen key features, effectively improves the distributed new energy power prediction accuracy in combination with scale-division modeling, and provides reliable data support for power grid dispatching.

Description

Technical Field

[0001] This invention relates to the field of renewable energy load forecasting technology, and in particular to a distributed renewable energy load forecasting method based on complete set empirical mode decomposition. Background Technology

[0002] With the rapid development of the distributed new energy industry, distributed new energy has transformed from a supplementary energy source to a core component of electricity growth, with its coverage and output scale continuously expanding at the distribution network level. However, the output characteristics of distributed new energy sources such as photovoltaics and wind power are greatly affected by meteorological conditions, exhibiting strong randomness, volatility, and intermittency. Simultaneously, with the large-scale integration of new loads such as electric vehicle charging piles, distributed energy storage, and smart electrical devices, the uncertainty of the distribution network's net load has further intensified. This not only leads to frequent problems such as bidirectional changes in distribution network power flow and voltage anomalies but also places higher demands on the real-time and accuracy of grid regulation. Factors affecting the accuracy of distributed new energy power prediction mainly include: a single-scenario focus and a lack of multi-regional collaborative perspective; existing research primarily focuses on centralized new energy or single-region scenarios, with less attention paid to collaborative prediction across multiple distribution network areas; insufficient feature utilization and inadequate key information mining; distributed energy power being affected by multiple dimensions; and weak model synergy and failure to adapt the optimal algorithm to the time scale. The distributed energy power sequence of a distribution area contains fluctuation characteristics at different time scales, including both long-term trend changes and short-term instantaneous fluctuations.

[0003] Therefore, improving the accuracy of distributed renewable energy power forecasting has become a core technical requirement for ensuring the stable operation of the new power system, promoting renewable energy consumption, and optimizing grid dispatch. This invention proposes a distributed renewable energy load forecasting method based on complete set empirical mode decomposition to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies in distributed renewable energy power prediction, such as single-scenario focus, insufficient feature utilization, and weak model synergy. The proposed method is a distributed renewable energy load prediction method based on complete set empirical mode decomposition, which improves the accuracy of distributed renewable energy power prediction, ensures the stable operation of the new power system, promotes renewable energy consumption, and optimizes grid dispatch.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A distributed renewable energy load forecasting method based on complete set empirical mode decomposition includes the following:

[0007] Through complete set empirical mode decomposition, the basic data is decomposed into multiple intrinsic mode functions and remainder terms.

[0008] Based on the decomposed intrinsic mode functions and remainder terms, an optimized feature set is constructed by improving the feature importance of the tree output according to the extreme gradient.

[0009] Based on the optimized feature set, long-term dependent subsequences are trained using a long short-term memory network, and short-term fluctuating subsequences are trained using an extreme learning machine to perform multi-scale modeling and obtain the load prediction values ​​of the subsequences.

[0010] The load prediction values ​​of the subsequences are superimposed to obtain the final power prediction result.

[0011] Furthermore, the basic data includes a time-series load dataset for a full year prior to the current time point for model training, and a validation load dataset measured at specified sites for complete prediction results.

[0012] Furthermore, the complete set empirical mode decomposition decomposes the basic data into multiple intrinsic mode functions and remainder terms, as follows:

[0013] By adding paired positive and negative white noise to suppress mode aliasing, positive and negative symmetric random white noise is added to the original power sequence of the time-series load dataset;

[0014] For each pair of noisy signals, empirical mode decomposition is performed to obtain the intrinsic mode function and the remainder term;

[0015] ;

[0016] in, , The first The first positive and negative noise decomposition under the second positive and negative noise decomposition One eigenmode function component; The number of times white noise is added. This represents the th result obtained after lumped average. Each intrinsic mode function component.

[0017] Furthermore, the empirical mode decomposition results are lumped and averaged using the following formula to eliminate noise interference:

[0018] ;

[0019] In the formula, This represents the remainder obtained after ensemble averaging. Indicates the first The remainder term obtained under subpositive noise decomposition; Indicates the first The remainder term obtained from the sub-negative noise decomposition.

[0020] Furthermore, based on the decomposed intrinsic mode functions and remainder terms, the optimized feature set is constructed according to the feature importance of the tree output based on extreme gradients, including:

[0021] The expression for constructing trend features is as follows:

[0022] ;

[0023] In the formula, This represents the power forecast value required for grid dispatch. For the number of trees, For tree structure space, Indicates the first The mapping relationship between weather conditions and power for each tree. The characteristics representing the intrinsic mode functions and remainder terms of the input sample;

[0024] Introduce the loss function as follows:

[0025] ;

[0026] In the formula, Indicates the loss of the sample. Indicates the number of samples;

[0027] Regularization terms used to control complexity: ;

[0028] Where T is the number of leaf nodes. Leaf weight, , This is the penalty coefficient;

[0029] The objective function is approximated using a second-order Taylor expansion, and the optimal split point is found using a greedy algorithm. The split gain is calculated using the following formula:

[0030] ;

[0031] in, , These are the sum of the first and second derivatives of the left subtree, respectively. , These are the sum of the first and second derivatives of the right subtree, respectively.

[0032] By plotting the curve of the prediction error of the validation load dataset as a function of the number of retained features K, the minimum K value that makes the model performance tend to stabilize is determined. The top K features in terms of importance are selected to generate an optimized feature set with trend characteristics for the transformer area.

[0033] Furthermore, the long-term dependent subsequences, based on the optimized feature set, are trained using a long short-term memory network, as follows:

[0034] The long-term dependency subsequences selected from the optimized feature set are normalized to the interval [0,1] using Min-Max, and represented as follows:

[0035] ;

[0036] in, Indicates time The normalized value of the long-term dependency class subsequence; Represents a long-term dependent subsequence at time [time]. The signal value, , These represent the maximum and minimum values ​​of the signal for the long-term dependent subsequence in the entire optimized feature set, respectively.

[0037] Furthermore, normalized trend prediction values ​​are generated through the output layer of the Long Short-Term Memory network:

[0038] ;

[0039] in, This represents the activation function. This represents the output weight vector. This indicates that the time step is hidden in the state. This indicates the output bias.

[0040] Furthermore, the short-term fluctuation subsequences are trained using an extreme learning machine to construct the output function expression of a single hidden layer feedforward neural network model, as follows:

[0041] ;

[0042] in, This represents the activation function. This represents the number of hidden layer nodes. The eigenmode functions and remainders of the input sample are represented by the following characteristics. , , These are the input weights of the activation function, the hidden layer bias, and the output weights, respectively. For the first The expected output of each sample, This is the index of the hidden layer node;

[0043] For output weights Solution:

[0044] The following was obtained analytically: ;

[0045] In the formula, Indicates generalized inverse, This indicates the expected output.

[0046] Furthermore, the load prediction values ​​of the subsequences are superimposed to obtain the final power prediction result, expressed as follows:

[0047] ;

[0048] In the formula, This indicates the final power prediction result obtained after fusion processing; This represents the optimization features in the optimization feature set. This represents the low-frequency reconstruction output of the Long Short-Term Memory network; This represents the high-frequency reconstructed output of the Extreme Learning Machine. and This is the weighted fusion coefficient.

[0049] Furthermore, based on the final power prediction results, the mean square error eRMSE and the maximum error eMAX are used as evaluation indicators to compare the deviation between the predicted values ​​and the actual values ​​of different distribution areas, and to analyze the fitting effect of long-term trends and short-term fluctuations.

[0050] The mean squared error eRMSE is expressed as follows:

[0051] ;

[0052] The maximum error eMAX is expressed by the following formula:

[0053] ;

[0054] In the above formula, This indicates the final power prediction result obtained after fusion processing. This represents the total number of samples in the test set. This represents the actual load power value. This indicates the highest historical load in the distribution radio station area.

[0055] Compared with existing technologies, the beneficial effects of this invention are as follows: The multi-model approach based on ensemble empirical mode decomposition, extreme gradient boosting trees, long short-term memory networks, and extreme learning machines provided by this invention offers excellent prediction accuracy. It can adaptively decompose power sequences and filter key features. Combined with multi-scale modeling, it effectively improves the accuracy of distributed renewable energy power prediction, providing reliable data support for power grid dispatch. Ensemble empirical mode decomposition can break down complex power sequences into subsequences at different time scales, specifically capturing various fluctuation characteristics and intuitively displaying the sources of fluctuations. This avoids the accuracy loss when a single model processes complex signals, ensuring the stability of prediction results. It has strong adaptability to multiple distribution areas and strong anti-interference capabilities. Extreme gradient boosting trees can simultaneously filter common and individual features across distribution areas, providing support for different geographical environments. Customized feature sets for load characteristics of distribution transformer areas avoid the accumulation of cross-regional errors caused by single-region modeling methods, meeting the needs of overall distribution network dispatch for collaborative prediction of multiple distribution transformer areas. The feature screening process can eliminate redundant information with weak correlation to power output, reduce noise interference caused by sudden weather changes such as cloud cover and short-term gusts, and output feature importance ranking, making it easier for dispatchers to understand the prediction logic and reducing operational difficulty and learning costs. The extreme learning machine algorithm used does not require iterative training. After randomly initializing the weights, it can quickly output the prediction results of short-term fluctuation subsequences through least squares solution. It has outstanding real-time performance and is suitable for real-time calculation scenarios of dynamic data at the edge. It meets the real-time dispatch requirements of distribution network prediction every 15 minutes without waiting for a complex calculation process, thus improving dispatch response efficiency.

[0056] Furthermore, this invention is highly practical. The input data consists of historical power, meteorological, and basic data of distribution networks that are readily available, eliminating the need for additional equipment to collect special data and reducing application costs. It also exhibits good adaptability to small-sample scenarios and supports transfer learning. For newly built distribution areas and scenarios with scarce historical data, it can achieve high-precision predictions without the need for extensive data training and without waiting for data accumulation periods. This accelerates the application of the model in newly added distribution areas, improves the accuracy of distributed renewable energy power prediction, ensures the stable operation of the new power system, promotes renewable energy consumption, and optimizes grid dispatch. Attached Figure Description

[0057] Figure 1 This is a flowchart of the distributed renewable energy load forecasting method based on complete set empirical mode decomposition in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the predictions of different models for Station 1 in this embodiment of the invention; Figure 3 This is a schematic diagram of different models predicting the performance of station No. 2 in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. These embodiments are merely preferred examples of this invention, used to aid in understanding the inventive concept, and do not constitute a limitation on the scope of protection.

[0059] The output characteristics of distributed renewable energy sources such as photovoltaics and wind power are greatly affected by meteorological conditions, exhibiting strong randomness, volatility, and intermittency. Photovoltaic output depends on solar radiation intensity and is easily affected by cloud cover and weather changes; wind power output is closely related to wind speed and direction, and short-term gusts and seasonal wind changes can cause drastic fluctuations in output. This embodiment proposes a distributed renewable energy load forecasting method based on complete set empirical mode decomposition. This forecasting method includes the following:

[0060] The basic data is preprocessed and decomposed into multiple intrinsic mode functions and remainder terms using Complete Set Empirical Mode Decomposition (CEEMD).

[0061] Based on the decomposed intrinsic mode functions and remainder terms, an optimized feature set is constructed according to the feature importance output by the extreme gradient boosting tree (XGBoost);

[0062] Based on the optimized feature set, long-term dependent subsequences are trained using a Long Short-Term Memory (LSTM) network with a fixed time step to capture long-term trend changes, while short-term fluctuating subsequences are trained using an Extreme Learning Machine (ELM) for multi-scale modeling to obtain the load prediction values ​​of the subsequences.

[0063] The load prediction values ​​of the subsequences are superimposed through the result fusion layer to obtain the final power prediction result.

[0064] Reference Figure 1 The underlying data in this example includes the time-series load dataset of the previous full year for model training, and the validation load dataset of the specified field measurements for complete prediction results.

[0065] The basic data is preprocessed to remove outliers and missing values, and then standardized. Next, it undergoes Complete Set (Complementary Set) Empirical Mode Decomposition (CEEMD) to decompose the basic data into multiple intrinsic mode functions (EMFs) and residual terms. Meteorological characteristics such as solar irradiance, cloud cover, temperature, humidity, atmospheric transparency, wind speed, wind direction, turbulence intensity, and air density, as well as historical power data, are decomposed. The original non-stationary and nonlinear complex power sequence is decomposed into several high-frequency and low-frequency EMFs with clearer physical meaning, narrower bandwidth, and separation of trends and fluctuations, along with residual terms. The specific operations are as follows:

[0066] By adding paired positive and negative white noise to suppress mode aliasing, positive and negative symmetric random white noise is added to the original power sequence of the time-series load dataset;

[0067] ;

[0068] in, and This represents the wind and solar power generation capacity after adding white noise. This represents the power generation of wind and solar power within a certain time period. and Indicates the first The white noise added next has the same amplitude but opposite sign.

[0069] For each pair of noisy signals and Empirical mode decomposition is performed to obtain the intrinsic mode functions and remainder terms.

[0070] ;

[0071] in, , The first The first positive and negative noise decomposition under the second positive and negative noise decomposition One eigenmode function component; The number of times white noise is added is used to add positive and negative symmetrical noise multiple times during the complete set of empirical mode CEEMD decomposition to eliminate noise interference and improve the stability and accuracy of the decomposition. This represents the th result obtained after lumped average. The intrinsic mode function components are one of the stationary subsequences obtained after decomposing the original power sequence, and are used for subsequent subscale modeling.

[0072] The following formula is used to perform a lumped average on the decomposition results to eliminate noise interference;

[0073] ;

[0074] This represents the remainder after lumped averaging, reflecting the remaining trend components after the original power sequence decomposition, and participating in subsequent modeling to capture long-period changes; Indicates the first The remainder term obtained under subpositive noise decomposition; Indicates the first The remainder term obtained from the sub-negative noise decomposition.

[0075] Decoupling slowly changing trends (such as power baselines caused by sunrise and sunset) from rapid random fluctuations (such as sudden changes in cloud cover) facilitates subsequent separate modeling or feature extraction. The original power sequence contains multiple stationary or weakly stationary subsequences, and each intrinsic mode function (IMF) is more easily fitted by Extreme Gradient Boosting (XGBoost). After predicting and reconstructing each IMF separately, the IMFs are fed into the XGBoost as additional input features to improve overall accuracy.

[0076] The nonlinear linear features and trend components of the decomposed subsequences are captured. Based on meteorological and historical power features, and using the intrinsic mode functions and remainders after decomposition, trend features are extracted according to the feature importance output by the Extreme Gradient Boosting Tree (XGBoost) algorithm after complete set (complementary set) empirical mode decomposition (CEEMD). These features are then used for training of Extreme Learning Machine (ELM) and Long Short-Term Memory (LSTM) networks. The specific process includes:

[0077] The expression for constructing the trend feature model is as follows:

[0078] ;

[0079] In the formula, This represents the power forecast value required for grid dispatch. For the number of trees, For tree structure space, Indicates the first The mapping relationship between weather conditions and power for each tree. The intrinsic mode functions and remainders of the input sample are represented by features (including meteorological, historical power, time, etc.).

[0080] To minimize the total loss and achieve a balance between fitting accuracy and model complexity, a loss function (including a regularization term) is introduced:

[0081] ;

[0082] In the formula, The loss of the sample is used to measure the actual power value. Compared with power prediction values The error between them reflects the accuracy of the model's predictions; This indicates the number of samples, that is, the total number of samples used in model training or computation.

[0083] Regularization terms used to control complexity: ;

[0084] Where T is the number of leaf nodes. Leaf weight, , This is the penalty coefficient;

[0085] Solution optimization:

[0086] If a feature frequently serves as a split point across multiple trees, it indicates a significant contribution to reducing power prediction errors. A second-order Taylor expansion is used to approximate the objective function, and a greedy algorithm is employed to find the optimal split point. The split gain is calculated using the following formula:

[0087] ;

[0088] in, , These are the sum of the first and second derivatives of the left subtree, respectively. , These are the sum of the first and second derivatives of the right subtree, respectively.

[0089] For each distribution substation, an XGBoost regression model is independently constructed using the method described above, and a pre-defined hyperparameter combination space is traversed using 5-fold cross-validation (the data is divided into 5 parts, which are used alternately as validation load datasets). The average feature importance calculated on each validation load dataset is used as the evaluation criterion to systematically analyze the contribution of different feature subsets to model performance. Furthermore, by plotting the prediction error (e.g., RMSE) of the validation load dataset as a function of the number of retained features K, the minimum value of K that makes the model performance tend to stabilize is determined, which is the optimal number of features for that distribution substation. Finally, the top K features by importance are selected to generate an optimized feature set with trend characteristics for that distribution substation, which is used for subsequent high-precision power prediction and scheduling decisions.

[0090] Extreme gradient boosting (XGBoost) can simultaneously filter common and individual features across distribution areas, customizing feature sets for distribution areas with different geographical environments and load characteristics. This avoids the accumulation of cross-distribution area errors caused by single distribution area modeling methods and meets the needs of overall distribution network dispatch for collaborative prediction of multiple distribution areas.

[0091] Based on the optimized feature set, long-term dependent subsequences are trained using a Long Short-Term Memory (LSTM) network with a fixed time step to capture long-term trend changes. The process is as follows:

[0092] The long-term dependency subsequences selected from the optimized feature set are normalized to the interval [0,1] using Min-Max, and represented as follows:

[0093] ;

[0094] in, Indicates time The normalized values ​​of long-term dependent class subsequences (decomposed intrinsic mode functions (IMF) and remainder terms (RE)). Represents a long-term dependent subsequence at time [time]. The signal value, , These represent the maximum and minimum values ​​of the signal in the long-time dependent subsequence of the entire optimized feature set, respectively.

[0095] Normalized trend prediction values ​​are generated through the output layer of a Long Short-Term Memory (LSTM) network:

[0096] ;

[0097] in, This represents the activation function. This represents the output weight vector. This indicates that the time step is hidden in the state. This indicates the output bias.

[0098] When training with a Long Short-Term Memory (LSTM) network, gradient vanishing or exploding can occur when the LSTM network processes long-term dependent subsequences, causing the model to be unable to learn long-term dependencies (such as yesterday's weather affecting today's power). By introducing memory units (cell states) and gating mechanisms, the long-term dependencies and dynamic changes of subsequences can be captured, enabling selective retention, updating and output of information, thereby effectively capturing long-term dependencies across time steps.

[0099] Input gate, forget gate, output gate, and cell state control the flow of information through gating mechanisms.

[0100] Forgotten Gate: ;

[0101] Input Gate: ;

[0102] Cell candidate state: ;

[0103] Cell status update: ;

[0104] Output gate: ;

[0105] In the above formula, For the Gate of Oblivion The output at each time step is used to control the proportion of information that needs to be forgotten in the cell state of the previous time step, adapt to the long-term dependence characteristics of the power sequence of multiple distribution areas in the distribution network, and filter historical information that is valuable for the current prediction. The weight matrix from the hidden layer to the forget gate in the forget gate reflects the degree of influence of the historical power hiding characteristics of the distribution network on the forget decision. for The hidden state at time -1 contains key characteristic information about the power of multiple distribution areas in the previous time. This is the weight matrix from the input layer to the forget gate in the forget gate, which reflects the role of the input characteristics of multiple distribution network areas (such as weather, installed capacity, etc.) on the forget decision at the current moment; for The input characteristics at any given time cover real-time power, weather and other influencing factors in multiple distribution network areas; This is the weight matrix from cell state to forget gate in the forget gate, used to combine historical information in cell state to assist forgetting decision-making; The cell state at time t-1 stores key state information of the power sequence of multiple distribution areas in the previous time step. The bias term of the forget gate is used to adjust the output of the forget gate to adapt to the power prediction deviation in the power distribution network scenario.

[0106] For input gates The output at each moment controls the proportion of new input information that needs to be retained in the cell state at the current moment. For the dynamic changes in the power of multiple distribution areas, it filters out features in the current input that are useful for prediction. This is the weight matrix from the hidden layer to the input gate, reflecting the influence of hidden layer features on the input information retention decision; The weight matrix from the input layer to the input gate in the input gate reflects the role of the current input features in the information retention decision; This is the weight matrix from cell state to input gate in the input gate, which combines historical cell state information to assist input decision-making; The input gate's bias term is used to adjust the input gate output to adapt to the power prediction scenario of the power distribution network.

[0107] For cell candidate states in The output at any given time is used to generate candidate update values ​​for cell states, capturing the potential changing trends of power sequences across multiple distribution areas in the power distribution network; This is the weight matrix from the hidden layer to the cell candidate state, reflecting the influence of hidden layer features on the cell candidate state; This is the weight matrix from the input layer to the cell candidate state, reflecting the effect of the current input features on the cell candidate state; This is the weight matrix from cell state to cell candidate state in the cell candidate state, which combines historical cell state information to assist in candidate state generation; The bias term for the candidate states of the cells is used to adjust the candidate state output to adapt to the power prediction of the distribution network.

[0108] for The constantly updated cell state is stored. The key state information of the power sequence of multiple distribution areas in real time provides long-term memory support for subsequent forecasting; Input gate in The output at any given moment controls the retention ratio of the current input information; This is the weight matrix from the hidden layer to the cell state in the cell state update, reflecting the influence of the hidden layer features on the cell state update; This is the weight matrix from the input layer to the cell state in the cell state update, reflecting the effect of the current input features on the cell state update; This is a bias term for cell state updates, adjusting the cell state update results to fit the power prediction of the distribution network.

[0109] for The hidden state of time contains Key characteristic information of power in multiple distribution network areas at any given time is used to output prediction results.

[0110] The output function expression of the short-term fluctuation subsequence is constructed by training an Extreme Learning Machine (ELM) to build a single hidden layer feedforward neural network model, as follows:

[0111] ;

[0112] in, This represents the activation function. This represents the number of hidden layer nodes. The eigenmode functions and remainders of the input sample are represented by the following characteristics. , , These are the input weights of the activation function, the hidden layer bias, and the output weights, respectively. For the first The expected output of the nth sample, in the power prediction scenario of a multi-area distribution network, represents the nth sample. The moment or the first The actual power values ​​of each transformer area are used to measure the accuracy of the model's prediction results; The index is used to distinguish different hidden layer nodes. Each node corresponds to a set of input weights, biases and output weights, which jointly participate in the fusion calculation of multi-source features (such as power distribution area, meteorology, etc.).

[0113] For output weights Solution:

[0114] The following was obtained analytically: ;

[0115] In the formula, It represents the generalized inverse. In existing technologies, numerical stability is usually achieved by using SVD (Singular Value Decomposition). This indicates the expected output.

[0116] It is important to note that for scenario adaptability, when making ultra-short-term predictions, dynamic data from the edge can be accessed for real-time calculation. For small-sample scenarios (such as newly built transformer substations), the improved sparrow search algorithm can be used to optimize parameters, and transfer learning can be used to adapt to transformer substations with scarce data.

[0117] The load prediction values ​​of the subsequences are superimposed through the result fusion layer to obtain the final power prediction result, as shown in the following expression:

[0118] ;

[0119] In the formula, This indicates the final power prediction result obtained after fusion processing; This represents the optimization features in the optimization feature set. This represents the low-frequency reconstruction output of the Long Short-Term Memory network; This represents the high-frequency reconstructed output of the Extreme Learning Machine. and This is the weighted fusion coefficient.

[0120] Using mean squared error (eRMSE) and maximum error (eMAX) as evaluation indicators, we compared the deviations between predicted and actual values ​​for different transformer areas to analyze the fitting effect of long-term trends and short-term fluctuations.

[0121] The mean squared error eRMSE is expressed as follows:

[0122] ;

[0123] The maximum error eMAX is expressed by the following formula:

[0124] ;

[0125] In the above formula, This represents the total number of samples in the test set. This represents the actual load power value. This indicates the highest historical load in the distribution radio station area.

[0126] The effectiveness of this solution will be further illustrated below with specific examples.

[0127] The load data in this example all come from a State Grid power supply company, mainly covering real-time load monitoring data of photovoltaic power stations and wind power stations within the company's jurisdiction; the weather data uses the meteorological observation data released by the local meteorological department of the experimental area during the same period to ensure the timeliness of the data.

[0128] The data is divided into three parts: First, the load data from 00:00 on January 1, 2024 to 23:45 on December 31, 2024 is set as the training set. This training set has a sampling period of 15 minutes and contains 96 continuous sampling points per day, forming a complete annual time-series load dataset for model parameter learning and feature fitting. Second, the load data from 00:00 on January 1, 2025 to 23:45 on February 28, 2025 is set as the prediction target. The prediction granularity is consistent with the sampling period of the training set, that is, the load prediction results are output every 15 minutes. Finally, the measured data from 10 designated stations provided by the company is used as the validation set of the model to verify the accuracy and reliability of the prediction effect.

[0129] To ensure the effectiveness of data decomposition, the key parameters for CEEMD in this example are set as follows:

[0130] The noise bias coefficient is set to 0.15 to control the intensity of white noise added to the original load data, balancing decomposition accuracy and anti-interference capability. The number of white noise additions is set to 40 times to weaken mode aliasing by superimposing white noise multiple times and improve the stability of the decomposition results. The maximum number of algorithm iterations is set to 200 times to ensure that each intrinsic mode function (IMF) can fully converge during the decomposition of the distributed new energy original power sequence in the basic data, meeting the input requirements of subsequent models.

[0131] After completing the CEEMD decomposition of the power sequences of each power distribution area and storing m IMF components and 1 RE remainder for later use, XGBoost feature selection is first constructed on the training set. Parameter combinations are traversed through 5-fold cross-validation, and the optimal K value for each power distribution area is determined based on the feature importance gain value of the validation set, generating a dedicated optimized feature set. Subsequently, scaled modeling is performed: long-term dependent subsequences are trained using LSTM with a fixed time step; short-term fluctuating subsequences are trained using ELM. The number of hidden layer neurons is determined by the error curve of the validation load dataset. Obtaining the error curve requires first dividing the experimental data into training and validation sets. Then, for the ELM model, after each adjustment of the number of hidden layer neurons, the model is trained using the training set and predicted using the validation set. Error indices such as relative mean square error (eRMSE) and maximum error (eMAX) are calculated. Finally, a curve is plotted with the number of hidden layer neurons as the x-axis and the error indices as the y-axis to determine the optimal number of hidden layer neurons, providing technical support for model accuracy. The Sigmoid activation function is chosen for rapid calculation of output weights. Finally, prediction and result verification are performed. The test set is input into the two trained models, and the predicted values ​​of the subsequences are obtained and fused and superimposed according to time steps to obtain the actual power prediction results. The relative mean square error (eRMSE) and the maximum error (eMAX) are used as evaluation indicators to compare the deviation between the predicted values ​​and the actual values ​​of different transformer areas. The fitting effect of LSTM on long-term trends and ELM on short-term fluctuations is analyzed, and the change in model accuracy before and after feature selection is verified.

[0132] Based on measured data from 10 designated sites, the results of Complete Ensemble Empirical Mode Decomposition (CEEMD) are as follows:

[0133] IMF1 to IMF4 are high-frequency components, and their waveforms exhibit significant high-frequency oscillation characteristics, with delicate and dense fluctuation patterns. Among them, IMF1 has the smallest amplitude but the highest oscillation frequency, accurately capturing the most subtle and fastest instantaneous fluctuation characteristics in the original signal; IMF3 maintains a high-frequency oscillation state with a slightly larger amplitude, together forming the core carrier of the high-frequency dynamic components in the original signal.

[0134] IMF5 to IMF8 belong to the mid-frequency components, with oscillation frequencies significantly lower than the previous four orders, and the graininess of the waveform tends to be smoother. Taking IMF5 as an example, its oscillation period is significantly longer, and the amplitude and fluctuation rhythm form a unique ratio, clearly reflecting the periodic variation law of the medium scale in the original signal, and it is a key transition carrier connecting high and low frequency components.

[0135] IMF9 to IMF10 are low-frequency components, and their waveforms exhibit long-period, low-frequency evolution characteristics. In particular, IMF10 has the smoothest overall fluctuation rhythm, with its oscillation period reaching the maximum value of the entire sequence. It directly demonstrates the most stable and persistent low-frequency trend component in the original signal and is a core indicator characterizing changes in the signal base.

[0136] For station 1, such as Figure 2 As shown, the fitting trends of predicted values ​​and actual values ​​(Real) for three models—the LightGBM gradient boosting tree model, the LinearRegRession model, and the proposed multi-model approach—are presented. It is clearly visible that the predicted curve of the proposed multi-model approach fits the curve representing the actual value better, while the predicted curve of the LightGBM gradient boosting tree model shows a more significant deviation from the actual value.

[0137] like Figure 3 As shown, the prediction performance of the three models is also demonstrated for station 2. The prediction curve of this scheme is closer to the actual value curve at each stage of load increase, peak, and decrease; in contrast, the prediction curve of the gradient boosting tree model (LightGBM) deviates more from the actual value, especially in the peak load region, where the deviation is more prominent.

[0138] Table 1 Comparison of prediction results indicators of different algorithms

[0139]

[0140] The calculation data shown in Table 1 clearly demonstrates that the relative mean squared error (eRMSE) and maximum error (eMAX) of this scheme significantly outperform the gradient boosting tree model (LightGBM) and the linear regression model (LinearRegRession) in prediction accuracy. Specifically, the eRMSE of this scheme is only 0.146 and the eMAX is 0.198, far lower than the 1.701 and 0.826 of the gradient boosting tree model, and also better than the 0.293 and 0.301 of the linear regression model. This fully reflects the outstanding advantages of this implementation scheme in error control and prediction stability in distributed renewable energy power prediction tasks.

[0141] Although the steps in the above embodiments are described in the above order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not need to be executed in such order. They can be executed simultaneously (in parallel) or in reverse order. These simple variations are all within the protection scope of this invention.

[0142] Those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the claims of this invention, any of the claimed embodiments can be used in any combination.

[0143] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, the word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed PC.

[0144] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it should be noted that the parts not covered in this invention are the same as or can be implemented using existing technology. It will be readily understood by those skilled in the art that the scope of protection of this invention is obviously not limited to these specific embodiments. Without departing from the principles of this invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of this invention.

Claims

1. A distributed renewable energy load forecasting method based on complete set empirical mode decomposition, characterized in that, include: Through complete set empirical mode decomposition, the basic data is decomposed into multiple intrinsic mode functions and remainder terms; Based on the decomposed intrinsic mode functions and remainder terms, an optimized feature set is constructed by improving the feature importance of the tree output according to the extreme gradient. Based on the optimized feature set, long-term dependent subsequences are trained using a long short-term memory network, and short-term fluctuating subsequences are trained using an extreme learning machine to perform multi-scale modeling and obtain the load prediction values ​​of the subsequences. The load prediction values ​​of the subsequences are superimposed to obtain the final power prediction result.

2. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 1, characterized in that, The basic data includes a time-series load dataset for a full year prior to the current time point used for model training, and a validation load dataset measured at specified sites for complete prediction results.

3. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 2, characterized in that, The complete set empirical mode decomposition decomposes the basic data into multiple intrinsic mode functions and remainder terms, as follows: By adding paired positive and negative white noise to suppress mode aliasing, positive and negative symmetric random white noise is added to the original power sequence of the time-series load dataset; For each pair of noisy signals, empirical mode decomposition is performed to obtain the intrinsic mode function and the remainder term; ； in, , The first The first positive and negative noise decomposition under the second positive and negative noise decomposition One eigenmode function component; The number of times white noise is added. This represents the first digit obtained after ensemble average. Each intrinsic mode function component.

4. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 3, characterized in that, The empirical mode decomposition results are lumped and averaged using the following formula to eliminate noise interference: ； In the formula, This represents the remainder obtained after ensemble averaging. Indicates the first The remainder term obtained under subpositive noise decomposition; Indicates the first The remainder term obtained from the sub-negative noise decomposition.

5. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 1, characterized in that, The optimized feature set is constructed based on the decomposed intrinsic mode functions and remainder terms, according to the feature importance of the boosting tree output by extreme gradients, including: The expression for constructing trend features is as follows: ； In the formula, This represents the power forecast value required for grid dispatch. For the number of trees, For tree structure space, Indicates the first The mapping relationship between weather conditions and power for each tree. The characteristics representing the intrinsic mode functions and remainder terms of the input sample; Introduce the loss function as follows: ； In the formula, Indicates the loss of the sample. Indicates the number of samples; Regularization terms used to control complexity: ; Where T is the number of leaf nodes. Leaf weights , This is the penalty coefficient; The objective function is approximated using a second-order Taylor expansion, and the optimal split point is found using a greedy algorithm. The split gain is calculated using the following formula: ； in, , These are the sum of the first and second derivatives of the left subtree, respectively. , These are the sum of the first and second derivatives of the right subtree, respectively. By plotting the curve of the prediction error of the validation load dataset as a function of the number of retained features K, the minimum K value that makes the model performance tend to stabilize is determined. The top K features in terms of importance are selected to generate an optimized feature set with trend characteristics for the transformer area.

6. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 1, characterized in that, The process of training a long-term dependent subsequence network based on an optimized feature set is as follows: The long-term dependency subsequences selected from the optimized feature set are normalized to the interval [0,1] using Min-Max, and represented as follows: ； in, Indicates time The normalized value of the long-term dependency class subsequence; Represents a long-term dependent subsequence at time [time]. The signal value, , These represent the maximum and minimum values ​​of the signal for the long-term dependent subsequence in the entire optimized feature set, respectively.

7. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 6, characterized in that, Normalized trend prediction values ​​are generated through the output layer of the Long Short-Term Memory network: ； in, This represents the activation function. This represents the output weight vector. This indicates that the time step is hidden in the state. This indicates the output bias.

8. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 1, characterized in that, The short-time fluctuation subsequences are trained using an extreme learning machine to construct the output function expression of a single hidden layer feedforward neural network model, as follows: ； in, This represents the activation function. This represents the number of hidden layer nodes. The eigenmode functions and remainders of the input sample are represented by the following characteristics. , , These are the input weights of the activation function, the hidden layer bias, and the output weights, respectively. For the first The expected output of each sample This is the index of the hidden layer node; For output weights Solution: The following was obtained analytically: ; In the formula, Indicates generalized inverse, This indicates the expected output.

9. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 1, characterized in that, The load prediction values ​​of the subsequences are superimposed to obtain the final power prediction result, as shown in the following expression: ； In the formula, This indicates the final power prediction result obtained after fusion processing; This represents the optimization features in the optimization feature set. This represents the low-frequency reconstruction output of the Long Short-Term Memory network; This represents the high-frequency reconstructed output of the Extreme Learning Machine. and This is the weighted fusion coefficient.

10. The distributed renewable energy load forecasting method based on complete set empirical mode decomposition according to claim 9, characterized in that, Based on the final power prediction results, the mean square error eRMSE and the maximum error eMAX are used as evaluation indicators to compare the deviation between the predicted values ​​and the actual values ​​of different distribution areas, and to analyze the fitting effect of long-term trends and short-term fluctuations. The mean squared error eRMSE is expressed as follows: ； The maximum error eMAX is expressed by the following formula: ； In the above formula, This indicates the final power prediction result obtained after fusion processing. This represents the total number of samples in the test set. This represents the actual load power value. This indicates the highest historical load in the distribution radio station area.

Images