Multi-element two-stage adaptive accumulation prediction method for regional integrated energy system

Abstract

The application provides a multi-element two-stage adaptive accumulation prediction method for a regional integrated energy system; firstly, a new preprocessing module based on ensemble learning is provided to preprocess original input data, provide a reliable data basis for TAPM, and determine key input variables of multi-energy load prediction; then, the training results of four predictors are adaptively integrated in the first stage of TAPM to enhance the generalization ability of the model, wherein a new collaborative atomic chaotic search algorithm is provided to train the predictor hyperparameters and intermediate data set of TAPM, so that the adaptive accumulation of the predictor in TAPM is realized; finally, peak load prediction correction in the power consumption peak period is carried out in the second stage of TAPM; the energy cascade utilization efficiency is improved, the regional integrated energy system is widely concerned, the multi-energy load prediction is very crucial for the planning of the regional integrated energy system, and effective decisions can be made by managers for the carbon peak target.

Description

Technical Field

[0001] This invention belongs to the field of load forecasting technology, and relates to multi-energy load preprocessing methods, heuristic optimization methods, integrated learning load forecasting methods, and especially multi-energy two-stage adaptive stacking forecasting methods for regional integrated energy systems. Background Technology

[0002] With rapid industrialization and rising living standards, the contradiction between energy supply and environmental issues has become increasingly prominent, making it imperative to promote energy structure transformation, improve energy utilization efficiency, and foster sustainable development. Integrated energy systems, as a major carrier of the energy internet, encompass various energy sources such as electricity, heat, and cooling, enabling deep coupling of different energy types and gradually becoming a major research direction in the energy field. Accelerating the construction of regional integrated energy systems helps improve the comprehensive utilization efficiency of various energy forms and contributes to environmental protection. Regional integrated energy systems utilize advanced physical information system technologies and management models to integrate multiple forms of energy, including electricity, heat, and cooling, within a certain spatial and temporal range, thereby systematically achieving lean operation and maintenance of diversified energy sources. With the promotion of regional integrated energy systems, multi-energy load forecasting has become a crucial prerequisite for lean operation and maintenance, ensuring that management departments can make accurate scheduling and intelligent maintenance decisions. The accuracy, flexibility, and efficiency of multi-energy load forecasting are of great significance for achieving supply and demand balance in regional integrated energy systems. However, the intermittent output, uneven distribution, and seasonal variations of renewable energy sources pose challenges to the reliable power supply of regional integrated energy systems and the load forecasting of multi-energy systems. Because the coupling relationships between different integrated energy subsystems are difficult to define, the influencing factors of multi-energy loads are more complex than those of traditional power distribution networks.

[0003] Due to the indispensable role of ultra-short-term load forecasting in power systems, related technological research is rapidly developing. Existing independent forecasting models can be broadly categorized into parametric algorithms and data-driven algorithms. However, uncertain meteorological conditions and social events introduce numerous uncertainties to the multi-energy load of regional integrated energy systems, and parametric algorithms struggle to address nonlinear problems and analyze complex influencing factors. Data-driven algorithms, such as Random Forest (RF), Support Vector Regression (SVR), and artificial neural networks, can rapidly predict multi-energy loads in ultra-short-term load forecasting. Tianjin University proposed a Wavelet Neural Network (WNN) model, which optimizes the ultra-short-term load forecasting model for regional integrated energy systems using improved Particle Swarm Optimization (IPSO) and Chaos Optimization Algorithm (COA). While a single forecasting model can rapidly predict multi-energy loads in certain scenarios, it may exhibit poor generalization ability and overfitting in other or more complex scenarios.

[0004] Employing ensemble learning to predict the multi-energy load of regional integrated energy systems has become an important approach to solving the aforementioned problems. Multiple data-driven models are integrated into an ensemble prediction framework to achieve better performance than any single model with complementary mechanisms. Appropriate ensemble strategies can provide better fitting performance and prediction accuracy. Mohammed proposed a Multivariate Ensemble Forecasting Framework (MEFF), which systematically combines three multi-predictors: an Elman neural network, a feedforward neural network, and a radial basis function neural network. These predictors are trained using global particle swarm optimization to improve their prediction accuracy within the MEFF. Some researchers have proposed a hybrid model that integrates Long Short-Term Memory (LSTM), encoder-decoder, and gradient boosting decision trees to improve the prediction accuracy of multi-energy load. Among numerous ensemble techniques, the Stacking Ensemble Framework (SEF) is renowned for its excellent robustness and prediction accuracy and is increasingly being applied to the field of multi-energy load prediction.

[0005] It is worth noting that existing literature has limited research on improving the generalization ability and training efficiency of stacked ensemble frameworks. Traditional SEF, when generating intermediate datasets for integrating the training results of all base predictors, suffers from interference from weak predictors. The weighted average of each predictor's output weakens the performance of strong predictors, and the correlation of each prediction model is ignored. Furthermore, simply stacking predictors in a single, disordered manner actually reduces the overall training speed of SEF, with limited improvement in accuracy. The resulting huge computational cost makes SEF difficult to apply to large-scale integrated energy systems. Meanwhile, the prediction error of multi-energy loads mainly originates from peak energy consumption periods, and the uncertainty of meteorological changes and the coupling effects between different energy forms make it difficult for most existing prediction methods to reduce the prediction error of peak loads. Furthermore, traditional ultra-short-term load forecasting methods mainly focus on forecasting only one type of load. Proposing a lightweight SEF that coordinates the forecasting of electricity, heat, and cooling loads for regional integrated energy systems has extremely high application value. Dynamically changing the SEF structure to adapt to the characteristics of the base forecasting model is a meaningful exploration of SEF. Peak load forecasting calibration based on SEF with strong generalization ability is an effective way to improve the accuracy of multi-energy load forecasting.

[0006] To fully exploit the potential of SEF (Search Engine Optimization), a suitable method is urgently needed to optimize the basis predictor weights and hyperparameter configurations. However, the search space for such problems grows exponentially, and traditional optimization methods often only produce suboptimal solutions. Gradient-free mechanisms allow metaheuristic algorithms to bypass the calculation of search space derivatives, thereby reducing computational costs and improving search performance. Population-based metaheuristic algorithms, such as Genetic Algorithm (GA), Gray Wolf Optimizer (GWO), and Particle Swarm Optimization (PSO), have been widely praised in hyperparameter optimization methods. Recently, Azizi proposed the Atomic Orbital Search (AOS) algorithm, a novel metaheuristic algorithm based on quantum atom models and quantum mechanics principles. AOS outperforms other metaheuristic algorithms in global convergence performance across multiple mathematical functions with different dimensions. However, AOS remains challenging in handling complex optimization problems, especially in dealing with the coupling relationships between different energy forms. Furthermore, AOS's photon numerical strategy and magnetic field simulation strategy cannot reasonably balance exploration and utilization and cannot provide a convincing scientific explanation. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and propose a multi-electrode, two-stage adaptive stacking prediction method for regional integrated energy systems. By deeply studying the coupling characteristics of multi-energy load data for regional integrated energy systems, the multi-energy load of regional integrated energy systems can be accurately predicted.

[0008] The present invention solves its practical problem by adopting the following technical solution:

[0009] A multi-stage adaptive stacking prediction method for regional integrated energy systems, wherein the adaptive stacking prediction method is based on a preprocessing module, a first adaptive stacking prediction module, and a second adaptive stacking prediction module.

[0010] The preprocessing module is used to preprocess, decompose, and recombine the original input data to generate clustered fusion data;

[0011] The first adaptive stacking prediction module is used to calculate the cold load, heat load, and electrical load from the clustered fusion data to generate load prediction data;

[0012] The first adaptive stacking prediction module is used to perform error estimation of peak cooling load, heating load, and electrical load from the load prediction data to generate load correction data; it includes the following steps:

[0013] Data preprocessing: Call the preprocessing module to preprocess the training error and influencing factors obtained in the first stage to meet the requirements of subsequent processes;

[0014] Variable dimensionality reduction: The preprocessing module is invoked, and the RF-RFE method is used to select representative influencing factors of peak load forecasting error;

[0015] Error estimation: Based on the preliminary load prediction results of the first adaptive stacking prediction module, the peak load range with large prediction errors can be identified; based on the data of the second adaptive stacking prediction module, the peak load period with prediction errors can be predicted; when retraining the second adaptive stacking prediction module, the original training labels are replaced with the preliminary training error of the first APM model to estimate the prediction error of multi-energy peak load on the test set.

[0016] Error correction: The prediction error on the test set is fed back to the first adaptive stacking prediction module to correct the prediction result. The Intrinsic Mode Function (IMF) fusion for each load type is linearly added. The sum of the preliminary multi-energy load prediction result calculated by the first adaptive stacking prediction module and the prediction error output by the second adaptive stacking prediction module is added together.

[0017] Furthermore, both the first adaptive stacking prediction module and the first adaptive stacking prediction module include three base predictors, one meta predictor, and an optimizer; wherein: the optimizer generates adaptive stacking intermediate parameters by performing a cooperative quantum chaotic search on the load prediction data and load correction data using an atomic chaotic search algorithm, including the following steps:

[0018] S1-1: Constructing a quantum-based atom search model;

[0019] S1-2: Determine the initial position of the candidate solution, calculate the initial degree of the candidate solution, randomly generate the number of electron shells, assign the candidate solutions to the electron shells, calculate the binding state, binding energy, and lowest energy level of the electron shell, and generate a random vector;

[0020] S1-3: The dynamic photon strategy determines whether electrons exhibit photon behavior. Some electrons absorb or undergo changes in their energy state, while others enter chaotic orbits for large-scale searching, thus completing the search for the globally optimal electron in the cooperative atom model.

[0021] Furthermore, the process of determining whether an electron exhibits photonic behavior through a dynamic photon strategy, where some electrons absorb energy or change their energy state, and others enter chaotic orbits for large-scale searching, thus completing the search for the globally optimal electron in the cooperative atom model, includes the following steps:

[0022] Step S1-3-1: The energy level of the i-th electron in the k-th electron shell of the dominant atom is greater than the binding state of the dominant atom in the k-th electron shell, i.e.: LO i,k ≥BE L,k The leading electron tends to emit photon energy based on multi-atom cooperation, and the position update equation in the cooperative mode is as follows:

[0023] LS i,k (λ+1)=LS i,k (λ)+α i (2β i ×LE-γ i ×BS L -σ i ×BS F ) / k

[0024]

[0025]

[0026]

[0027] In the formula: L Si,k (λ) and L Si,k (λ+1) is the λth and (λ+1th)th iteration position of the i-th electron in the k-th shell of the dominant atom, σ i It is a vector randomly generated in the range (0,1), BS L q represents the binding state of the dominant atom. L It is the number of electrons inside the dominant atom, BS F q represents the combination state of all dependent atoms. F It is the number of electrons in the subordinate atom;

[0028] Otherwise, electrons tend to absorb photons based on multi-atom cooperative absorption, and their cooperative position update equation is as follows:

[0029] LS i,k (λ+1)=LS i,k (λ)+α i (2β i ×LE L,k -γ i ×BS L,k -σ i ×BS F,k )

[0030]

[0031]

[0032]

[0033] In the formula: d L and d F The total number of electrons in the dominant and subordinate atoms of the kth hypothetical layer;

[0034] Step S1-3-2: The process of electrons entering chaotic orbits:

[0035] The chaotic space mapping is established using the following formula:

[0036]

[0037] In the formula: s ij (λ) is the j-th decision variable of the i-th candidate solution in the λ-th iteration; s ij,max and s ij,min These are the upper and lower bounds of the j-th decision variable, respectively; assuming S i (λ)=[s i1 (λ),s i2 (λ),...,s ij (λ),...,s im [(λ)] is the position vector of the i-th electron in the chaotic orbit at the λ-th iteration; each dimension s in Si(λ) ij (λ) maps to the chaotic variable c ij (λ)(1≥c ij (λ)≥0);

[0038] The following iterative logical equation is used for chaotic local search:

[0039] c ij (λ+1)=μc ij (λ)(1-c ij (λ))

[0040] In the formula: μ is the control coefficient. When μ is 4, the chaotic dynamics are characterized by iterative logic equations. Through a sufficient number of iterations, the chaotic orbit of an electron can traverse the entire search space.

[0041] The chaotic variable c ij (k) Mapping to the original solution space establishes the following solution space mapping:

[0042] s ij (λ+1)=(s ij,max -s ij,min )c ij (λ+1)+s ij,max

[0043] Step S1-3-3: Dynamic photon search is performed to meet different energy level optimization requirements by dynamically adjusting PR, including:

[0044] (1) When the fitness of electron i in the kth iteration is O i (k)≥historical average fitness of electron i O i,ave At this point, the electron under study is far from the lowest energy level, and the photon rate PR can dynamically increase to guide the electron into a chaotic orbit, which can be expressed as:

[0045] PR = (PR max -PR min ) / 2+r PR (PR max -PR min ) / 2

[0046] In the formula PR max and PR min These are the maximum and minimum photon rates, r PR These are random coefficients in [0,1].

[0047] (2) When O i (k) <O i,av At this point, the electron under study orbits the electron with the lowest energy level. The electron should then undergo a more refined local search as the iterations proceed. The nonlinear decrease in the photon rate PR can be expressed as:

[0048] PR = PR max -λ(PR max -PR min ) / λ max

[0049] In the formula λ max It represents the maximum number of iterations.

[0050] Furthermore, the training process of the base predictor by the first adaptive stacking prediction module and the second adaptive stacking prediction module is as follows:

[0051] The original training data is divided into K folds; three base predictors are trained on the K-1 fold data to predict the remaining one fold data and the test data. Each fold of the training set to be predicted corresponds to an independent prediction model; new single-fold training data and a new test set are generated from the prediction results.

[0052] By repeatedly performing this iteration, each base predictor can obtain predictions for 5 sets of new training data and 5 sets of new test data.

[0053] After all base predictors have completed training and prediction, a meta predictor is constructed by concatenating the prediction results from the 5-fold training dataset.

[0054] The new training set for the detector can be represented as follows:

[0055] TRS i =[TRP i (1) TRP i (2) TRP i (3) TRP i (4) TRP i (5)] T

[0056] Where: TRS i TRP represents the new training set generated by the i-th base predictor. i (j) represents the new j-th fold training data predicted by the i-th base predictor;

[0057] A new test set for a base predictor is obtained by weighting the five sets of predicted test data, which can be represented as:

[0058]

[0059] Where: TES i TEP represents the new test set generated by the i-th base predictor. i (j) represents the j-th new test set predicted by the i-th base predictor, W i (j) represents the weight of the j-th new test set of the i-th base prediction; three new training and test datasets containing RF, SVR and LSTM features are transferred to the meta-predictor prediction module for integration.

[0060] Furthermore, the first adaptive stacking prediction module and the second adaptive stacking prediction module perform the following prediction process for the meta-predictor:

[0061] The three new training datasets obtained from the training of the base predictor are concatenated into a multidimensional training set, and the dimensions of the training set are merged.

[0062] The original training labels and the merged training set are used for training with an adaptive boosting method.

[0063] The new test dataset obtained during the training of the base predictor is used to predict future multi-energy load data.

[0064] Beneficial effects

[0065] 1. This invention proposes a novel preprocessing module based on ensemble learning to preprocess the original input data, providing a reliable data foundation for the multivariate two-stage adaptive stacking prediction method and determining the key input variables for multi-energy load prediction.

[0066] 2. In the first stage of the multivariate two-stage adaptive stacking prediction method, the training results of four predictors are adaptively integrated to enhance the generalization ability of the model. A novel cooperative atomic chaotic search algorithm is proposed to train the predictor hyperparameters and intermediate dataset of the multivariate two-stage adaptive stacking prediction method to achieve adaptive stacking of predictors in the multivariate two-stage adaptive stacking prediction method.

[0067] 3. In the second stage of the multivariate two-stage adaptive stacking prediction method, the peak load prediction during peak electricity consumption periods is corrected. The proposed prediction method has higher prediction accuracy, generalization ability and stability than other benchmark prediction models. Attached Figure Description

[0068] Figure 1 This is a flowchart of a novel cooperative atomic chaotic search algorithm of the present invention;

[0069] Figure 2 This is a schematic diagram of a quantum-based atomic model according to the present invention.

[0070] Figure 3 This is a diagram of a multivariate adaptive stacking prediction model architecture of the present invention.

[0071] Figure 4 This is a flowchart illustrating the implementation of a multivariate two-stage adaptive stacking prediction method of the present invention. Detailed Implementation

[0072] The following is in conjunction with the appendix Figure 1-4 The embodiments of the present invention will be further described in detail below:

[0073] This invention provides a multi-element two-stage adaptive stacking prediction method (TAPM) for regional integrated energy systems, the flowchart of which corresponds to Figure 4 This includes: a novel preprocessing module based on ensemble learning, used to preprocess the raw input data, corresponding to... Figure 4The system includes an ensemble learning-based preprocessing module; a multivariate adaptive stacked prediction model (APM) that integrates three base predictors and one meta predictor to achieve ultra-short-term predictions of cooling load, heating load, and electrical load; and a novel cooperative atomic chaos search (CACS) algorithm for training the predictor hyperparameters and intermediate datasets of the APM. The first stage of the TAPM uses the constructed APM model to adaptively integrate the training results of the four predictors, corresponding to… Figure 4 The load forecasting module; the second stage of the TAPM uses the constructed APM model to correct the peak load forecast during peak electricity consumption periods, corresponding to... Figure 4 Error correction module.

[0074] In this embodiment, the novel preprocessing module based on ensemble learning includes a data reconstruction module and a variable dimensionality reduction module; the data reconstruction module is used to preprocess, decompose, and reorganize multi-energy loads; it includes the following steps:

[0075] (1) Data preprocessing

[0076] In fact, due to unplanned failures and regular equipment maintenance, the loss of multi-energy load data is common, necessitating preprocessing of the TAPM input data. Trigonometric interpolation, particularly suitable for interpolating periodic functions, is applied to fill in empty data in both the training and test sets. Furthermore, due to equipment failures, noise interference, communication delays, etc., measured data may deviate from actual conditions, and anomalous data significantly reduces the training effectiveness of the predictor in multi-energy load forecasting. Therefore, when state estimation is inconvenient, this invention employs an isolated forest approach to remove anomalous historical load data from regional integrated energy systems, as an anomaly detection method suitable for continuous numerical data. By randomly partitioning the characteristics of historical load data, sub-forest anomaly detectors are formed, and a basic forest anomaly detector is constructed. Data that is easily separated is considered anomalous and is modified.

[0077] (2) Data decomposition

[0078] Decomposing multi-energy loads into discrete components and predicting each component separately is an effective method to improve the load forecasting accuracy of regional integrated energy systems. As a signal processing method applicable to non-stationary and nonlinear signals, VMD overcomes the bandwidth limitations of wavelet methods and has more mathematical theoretical basis than empirical mode decomposition. Its goal is to decompose multi-energy loads into discrete components called intrinsic mode functions, each corresponding to a different frequency and residual.

[0079] (3) Clustering fusion

[0080] While predicting each intrinsic mode function (IMF) individually can improve prediction accuracy, the increase in output variables of the TAPM model slows down training speed and affects data updates for the prediction model. Therefore, this invention uses the SOM clustering algorithm to reorganize multiple IMFs obtained from VMD decomposition into different classes, making the regularity of the same group of IMFs more similar. Based on the cluster labels, the same group of IMFs is linearly added to form several new IMF sequences, reducing the amount of integrated data and highlighting the regularity of the IMF sequences. Finally, these reconstructed IMF sequences are used as the output variables of TAPM for prediction. The variable dimensionality reduction module is used to remove irrelevant interference information and select input variables that are highly correlated with the output variables. Since the random forest method has a built-in feature evaluation mechanism, it can measure the contribution of each input variable to the prediction result, thus eliminating insignificant features. To reduce redundant interference information between features, this invention uses a recursive feature elimination method based on random forest to reduce the dimensionality of input variables and select highly correlated input variables. It uses a random forest predictor to evaluate variable importance, eliminating the least correlated variables through multiple iterations, and finally retaining the combination of input variables with the highest prediction accuracy.

[0081] A multi-element, two-stage adaptive stacking prediction method for regional integrated energy systems, corresponding to Figure 4 The load forecasting module and error correction module include the following steps:

[0082] S1: A multivariate adaptive stacking prediction model (APM) is proposed, which integrates three base predictors and one meta predictor. It is the key sub-model for constructing TAPM and can achieve ultra-short-term prediction of multi-energy loads.

[0083] S2: A novel cooperative atomic chaotic search algorithm (CACS) is proposed. A quantum-based atomic model is established to perform cooperative quantum chaotic search to find the global optimal solution. The four predictors of APM are adaptively stacked through CACS to improve the prediction accuracy of the model.

[0084] S3: Proposes the first-stage multivariate adaptive stacking prediction method (TAPM first stage), constructs the first APM sub-model through S1 and S2, and makes preliminary predictions of the cooling load, heating load and electrical load of the regional integrated energy system based on historical data.

[0085] S4: A second-stage multivariate adaptive stacking prediction method (TAPM second stage) is proposed. A second APM sub-model is constructed through S1 and S2. The prediction error of S3 is used to estimate the prediction error of peak cooling load, heating load and electrical load, thereby correcting the multi-energy load prediction result of S3.

[0086] In this embodiment, the specific steps of S1 include:

[0087] S1.1: Select appropriate base predictors and meta predictors to construct the first multivariate adaptive stacked prediction model (APM). After processing by the preprocessing module based on ensemble learning, the future multi-energy loads in the regional integrated energy system can be initially predicted by the first APM model. In the first stage of TAPM, the training results of the four predictors are adaptively integrated.

[0088] In this embodiment, the specific steps of step S1.1 include:

[0089] TAPM comprises two independent APM models: one for load forecasting and the other for error correction. After processing by an ensemble learning-based preprocessing module, future multi-energy loads in the regional integrated energy system can be initially predicted using the first APM model. Although the prediction accuracy of APM is improved compared to a single predictor, the prediction error is still concentrated near the peak of multi-energy loads. Therefore, the initial prediction results and training errors of the first APM model are transferred to the second stage of TAPM. As a key sub-model of the multivariate two-stage adaptive stacked prediction method (TAPM), the essence of APM is to obtain initial prediction information through base predictors, dynamically generate new datasets based on the performance of each base predictor, and generalize and integrate the current prediction information to obtain more accurate prediction results than when running a single predictor. The logical flow of APM is as follows: Figure 3 As shown, the specific information is as follows:

[0090] (1) Predictor selection

[0091] In ensemble learning, it is necessary to select prediction methods with significant differences as base predictors. Different base predictors can predict multi-energy loads from various data spaces and structures during training. Then, the meta-predictor constructs a corresponding stacked model based on the relationship between the performance of different base predictors. This invention selects a combination of three base predictors with the best stacked prediction performance and one meta-predictor. Random Forest (RF) is an ensemble method composed of decision trees, where each decision tree depends on the value of independently sampled random vectors, and all decision trees have the same distribution. According to the law of large numbers, RF is less prone to overfitting when making predictions, so it is used as one of the base predictors in TAPM. Support Vector Machine (SVM) contains supervised machine learning models for classification and regression, and is widely considered the best supervised learning algorithm for solving regression and classification problems. Support Vector Regression (SVR) can support quantitative responses to input variables and predict unknown information. Long Short-Term Memory (LSTM) Neural Network is a special type of Recurrent Neural Network (RNN) that can learn long-term dependency information and is suitable for predicting large events with long intervals and delays in time series. Adaptive Boosting (AdaBoost) is a weighted weak classifier ensemble method based on boosting techniques. Its core idea is to train different weak classifiers on the same training set and then integrate these weak classifiers to form a strong classifier. Because AdaBoost can reuse different weighted combinations of the same training samples, it has lower requirements for the amount of training data. Its strong generalization ability allows AdaBoost to serve as the meta-predictor in TAPM, used to integrate the intermediate datasets generated by the three base predictors mentioned above.

[0092] (2) Base Predictor Training. To avoid overfitting, the intermediate dataset for APM is generated based on K-fold cross-validation. First, the original training data is divided into K folds (Training K); then, RF, SVR, and LSTM are trained on the K-1 fold data to predict the remaining one-fold data (Predict K for the new training set) and the test data (Predict K for the new test set). Each fold of the training set to be predicted corresponds to an independent prediction model (Model K); therefore, new single-fold training data and a new test set can be generated from the prediction results. By repeatedly repeating this iteration, each base predictor can obtain prediction values ​​for 5 folds of new training data and 5 sets of new test data. After all base predictors have completed training and prediction, a new training set for a meta-predictor is concatenated from the prediction results of the 5-fold training dataset, which can be represented as follows:

[0093] TRS i =[TRP i(1) TRP i (2) TRP i (3) TRP i (4) TRP i (5)] T

[0094] In the formula, TRS i TRP represents the new training set generated by the i-th base predictor. i (j) represents the new j-th fold training data predicted by the i-th base predictor. A new test set for a base predictor is obtained by weighting the five sets of predicted test data, which can be represented as:

[0095]

[0096] In the formula TES i TEP represents the new test set generated by the i-th base predictor. i (j) represents the j-th new test set predicted by the i-th base predictor, W i (j) represents the weight of the j-th new test set predicted by the i-th base predictor. The weights of the new test set should be adaptively adjusted considering the characteristics and performance of different base predictors.

[0097] Therefore, three new training and testing datasets containing RF, SVR, and LSTM features were transferred to the meta-predictor prediction module for integration.

[0098] (3) Meta-predictor prediction. The primary task of meta-predictor prediction is to summarize the knowledge provided by the base predictors trained on the base predictors and obtain preliminary prediction results through a meta-predictor that is less prone to overfitting. First, the three new training datasets obtained from the training of the base predictors are concatenated into a multi-dimensional training set. The dimensionality of the merged training set increases threefold.

[0099] Next, the original training labels and the merged training set are fed into the Adaptive Boosting (AdaBoost) method for training. Finally, the AdaBoost model uses the new test dataset obtained during the training of the base predictors to predict future multi-energy loads.

[0100] In this embodiment, the overall process of the Cooperative Atom Chaotic Search Algorithm (CACS) is as follows: Figure 1 As shown, the specific steps of S2 include:

[0101] S2.1: Establish a quantum-based atomic model, conduct cooperative quantum chaos search, and find the global optimal solution;

[0102] S2.2: Adaptively stacking APM using CACS to train the predictor hyperparameters and intermediate datasets for APM;

[0103] In this embodiment, step S2.1 specifically includes the following steps:

[0104] A quantum-based atomic search model is constructed, defining the solution space of the problem. Its core idea is to configure the absorption or emission of electron density and photon energy based on quantum-based atomic theory. A schematic diagram of the quantum-based atomic model is shown below. Figure 2 As shown, electrons are distributed in various hypothetical electron shells L. The electron cloud around the atomic nucleus is used as the search space, and each electron around the nucleus is considered a candidate solution. The electron position can be defined by decision variables, and the energy state of each electron serves as the objective function value for each candidate solution. Electrons with lower energy levels represent candidate solutions with better objective function values. These shells are used to delineate the search space of the AOS and simulate the wave-like behavior of electrons. The radius R of the shells indicates their distribution around the nucleus. The shell with the smallest radius represents the nucleus shell L0, and the shell with the largest radius L0 represents the nucleus shell L0. i The layers represent the first to the nth electron layers L. n The electron position can be determined using the electron probability density based on the probability density function (PDF). The candidate solution with the optimal objective function value represents the lowest energy level L in the atomic nucleus. E The electronic position.

[0105] Suppose the candidate solutions are as follows:

[0106]

[0107] In the formula, S represents the electrons surrounding the atomic nucleus. i It is the position of the i-th candidate solution in the atom, s ij Let be the j-th decision variable of the i-th candidate solution, q be the number of electrons in the electron cloud, m be the dimension of the electron position, and the initial random distribution of electrons is as follows:

[0108]

[0109] In the formula S ij,0 S represents the initial position of the candidate solution. ij,min S represents the minimum boundary of the j-th decision variable of the i-th candidate solution. ij,max R0 is the maximum boundary of the j-th decision variable in the i-th candidate solution, and R0 is a random number in the range [0,1]. The energy levels of different electrons are represented as follows:

[0110] O = [O1 O2…O] i …O q ] T i = 1, 2, ..., q

[0111] In the formula, O is a vector containing the objective function values ​​of all electrons. i Let be the energy level of the i-th electron. Based on the electron positions determined by the PDF, each electron shell contains partial candidate solutions. The positions and energy levels of electrons in different electron shells can be expressed as:

[0112]

[0113]

[0114] In the formula S k It is an electron in the kth electron shell, S i,k It is the i-th electron in the k-th electron shell, O k Here, n is the energy level of an electron in the k-th electron shell, n is the maximum number of electron shells, and d is the total number of electrons in the k-th electron shell. k It is the energy level of the electron in the k-th electron shell, O i,k S represents i,k The energy level.

[0115] The binding state BS of the k-th electron shell k and binding energy BE k The calculation is as follows:

[0116]

[0117]

[0118] The binding states BS and binding energies BE of atoms are calculated as follows:

[0119]

[0120]

[0121] A dynamic photon strategy based on different energy level optimization requirements is proposed. By dynamically adjusting the photon resonator (PR), strong electrons with better target values ​​tend to perform localized refined searches, while weak electrons with worse target values ​​tend to make large-scale modifications to explore a wider space, such as entering chaotic orbits. Assume O... i (λ) and O i,ave Let be the objective function value of the i-th electron at the λ-th iteration and the average objective function value of the i-th electron, respectively. The dynamic photon strategy is as follows:

[0122] (1)When O i (k)≥O i,ave At this point, the electron under study is far from the lowest energy level, and it needs to undergo significant modifications to explore a wider energy space. The photon rate PR can be dynamically increased to guide the electron into a chaotic orbit, which can be expressed as:

[0123] PR = (PR max -PR min ) / 2+r PR (PR max -PR min ) / 2

[0124] In the formula PR max and PR min These are the maximum and minimum photon rates, r PR It is a random coefficient in [0,1].

[0125] (2)When O i (k) <O i,ave At this point, the electron under study orbits the electron with the lowest energy level. The electron should then undergo a more refined local search as the iterations proceed. The nonlinear decrease in the photon rate PR can be expressed as:

[0126] PR = PR max -λ(PR max -PR min ) / λ max

[0127] In the formula λ max It represents the maximum number of iterations.

[0128] To maintain a reasonable balance between exploration and utilization, a cooperative atom model is proposed to enhance the electron optimization capability by integrating advanced knowledge from multiple atoms. It is assumed that an atom group consists of a dominant atom and multiple subordinate atoms, where the atom containing the lowest energy level electron is defined as the dominant atom, and the other atoms are considered subordinate atoms.

[0129] For dependent atoms, the electronic energy levels in each electron shell are related to BE. k A comparison is made to determine whether a photon is emitted or absorbed. The probability of a photon interacting with an electron depends on the randomly generated number φ and the photon rate PR. If φ ≥ PR and O i,k ≥BE k Then, the electron tends to emit a photon with a certain energy to reach the energy LE and state BS of the atom, which can be expressed as:

[0130]

[0131] In the formula S i,k (λ) and S i+1,k (λ+1) is the λth and (λ+1th)th iteration position of the i-th electron in the k-th layer.

[0132] α i β i and γ iIt is a vector randomly generated in the range (0,1) and is used to quantify the photon energy to be exchanged.

[0133] If φ≥PR and O i,k <BE k Electrons tend to absorb photons with a certain energy to reach state BE. k The lowest energy level of the electron in the k-th electron shell is LE. k It is represented as:

[0134]

[0135] In the formula LE k This represents a candidate solution with the optimal objective function value in the k-th electron shell.

[0136] Each electron of the dominant atom updates its position based on its own and its subordinate atoms' historical best positions, assuming LO i,k BE represents the energy level of the i-th electron in the k-th electron shell of the dominant atom. L,k This represents the binding state of the dominant atom in the k-th electron shell. If φ ≥ PR and LO i,k ≥BE L,k Then the leading electron tends to emit photon energy based on multi-atom cooperation, and the position update equation in the cooperative mode is as follows:

[0137] LS i,k (λ+1)=LS i,k (λ)+α i (2β i ×LE-γ i ×BS L -σ i ×BS F ) / k

[0138]

[0139]

[0140]

[0141] In the formula L Si,k (λ) and L Si,k (λ+1) is the λth and (λ+1th)th iteration position of the i-th electron in the k-th shell of the dominant atom, σ i It is a vector randomly generated in the range (0,1), BS L q represents the binding state of the dominant atom. L It is the number of electrons inside the dominant atom, BS F q represents the combination state of all dependent atoms. Fis the number of electrons within the subordinate atom. If φ≥PR and LO i,k <BE L,k , the electrons tend to absorb photons based on multi-atom cooperation, and its collaborative position update equation is as follows:

[0142] LS i,k (λ + 1)=LS i,k (λ)+α i (2β i ×LE L,k -γ i ×BS L,k -σ i ×BS F,k )

[0143]

[0144]

[0145]

[0146] In the formula, d L and d F are the total number of electrons of the dominant atom and the subordinate atom in the k-th imaginary layer.

[0147] If φ < PR, the electrons will enter the chaotic orbit to find the optimal solution. The original AOS algorithm simulates the influence of the magnetic field on electrons by randomly generating the vector r i . However, adding a random vector is essentially a simple random search, with low efficiency and a lack of scientific explanation, resulting in the waste of previously accumulated search experience. Given its ease of implementation and the ability to jump out of local optima, chaos theory has been applied to the field of optimization. The seemingly disordered process of chaotic variables has an inherent regularity. It uses the randomness, ergodicity, and regularity of chaotic variables to find the global optimum value. Therefore, by using the chaotic orbit c i to replace the atomic orbit r i , the blindness of random search can be effectively reduced, helping the electrons to get rid of the local optimum. The chaotic orbit of electrons can be expressed as follows:

[0148] (1) Chaotic space mapping. Assume S i (λ)=[s i1 (λ), s i2 (λ),..., s ij (λ),..., s im (λ)] is the position vector of the i-th electron in the chaotic orbit at the λ-th iteration. Each dimension s ij (λ) in Si(λ) is mapped to the chaotic variable c ij (λ)(1≥c ij (λ)≥0), which is expressed as follows:

[0149]

[0150] In the formula s ij (λ) is the j-th decision variable of the i-th candidate solution in the λ-th iteration. ij,max and s ij,min These are the upper and lower bounds of the j-th decision variable, respectively.

[0151] (2) Chaotic local search. The iterative logic equations used are as follows:

[0152] c ij (λ+1)=μc ij (λ)(1-c ij (λ))

[0153] In the formula, μ is the control coefficient. When μ is 4, the chaotic dynamics are characterized by iterative logic equations. Through a sufficient number of iterations, the chaotic orbit of an electron can traverse the entire search space, exhibiting pseudo-randomness, ergodicity, and irregularity.

[0154] (3) Solve the space mapping. Chaotic variable c ij (k) can be mapped to the original solution space as follows:

[0155] s ij (λ+1)=(s ij,max -s ij,min )c ij (λ+1)+s ij,max

[0156] In this embodiment, step S2.2 specifically includes the following steps:

[0157] When more predictors are integrated into APM, hyperparameter settings become very complex. Proper hyperparameter configuration can effectively improve the accuracy and computational efficiency of APM. Furthermore, the weight W... i The determination of (j) remains a controversial issue. Most existing studies only average the training results of each base predictor, without exploring W. i (j) is a dynamic change strategy. Since the features of each base prediction model and their inherent correlations are ignored, simply stacking the number of base predictors may result in excessively low computational speed and limited accuracy improvement. To achieve adaptive stacking in different scenarios and improve the generalization ability of the proposed framework, W... i (j) should be correlated with the predictive performance of different base prediction models in APM. Therefore, the selection... Figure 3Using the hyperparameters of the four predictor variables (RF, SVR, LSTM, and AdaBoost) and 15 intermediate weights as decision variables, and with prediction accuracy as the optimization objective, CACS is proposed as an adaptive stacking optimization method for APM. Even though the optimization results of each CACS run are not exactly the same, the differences in hyperparameters among each independent base prediction model can significantly improve the generalization ability of APM.

[0158] In this embodiment, step S3 specifically includes the following steps:

[0159] Due to the irregular electricity consumption behavior of users and the high penetration rate of distributed power sources, current forecasting methods struggle to eliminate prediction errors for peak multi-energy loads. Since APM (Advanced Performance Management) enhances generalization ability, peak prediction error correction based on APM is an effective method for addressing this problem. Therefore, this invention proposes a two-stage APM forecasting model for multi-energy load forecasting and peak error correction. In the first stage of TAPM, multi-energy loads are initially predicted by an independent APM model. Based on the results of the initial load forecast (first stage of TAPM), the range of peak loads with significant prediction errors is identified. Combined with the proposed ensemble learning-based preprocessing module, Figure 4 The implementation process of TAPM is described below.

[0160] In this embodiment, step S4 specifically includes the following steps:

[0161] In the second phase of TAPM, another independent APM model effectively mitigates peak load prediction errors due to its strong generalization ability. When training the second APM model, the original training labels are replaced with the initial training error of the first APM model to estimate the prediction error of multi-energy peak loads on the test set. Peak load prediction corrections for peak electricity consumption periods are performed in the second phase of TAPM. The prediction errors on the test set are fed back to the first APM model to correct the prediction results. The intrinsic mode functions (IMFs) of each load type are linearly added. The final prediction result of TAPM is the sum of the initial multi-energy load prediction results calculated by the first APM model and the prediction errors estimated by the second APM model. Based on the initial load prediction results (first phase of TAPM), the peak load range with larger prediction errors can be identified. The data range of the second APM model only includes peak load periods with larger prediction errors. When training the second APM model, the original training labels are replaced with the initial training error of the first APM model to estimate the prediction error of multi-energy peak loads on the test set.

[0162] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A multi-element, two-stage adaptive stacking prediction method for regional integrated energy systems, characterized by: The adaptive stacking prediction method is based on a preprocessing module, a first adaptive stacking prediction module, and a second adaptive stacking prediction module. The preprocessing module is used to preprocess, decompose, and recombine the original input data to generate clustered fusion data; The first adaptive stacking prediction module is used to calculate the cold load, heat load, and electrical load from the clustered fusion data to generate load prediction data; The first adaptive stacking prediction module is used to perform error estimation of peak cooling load, heating load and electrical load on the load prediction data to generate load correction data; Includes the following steps: Data preprocessing: Call the preprocessing module to preprocess the training error and influencing factors obtained in the first stage to meet the requirements of subsequent processes; Variable dimensionality reduction: The preprocessing module is invoked, and the RF-RFE method is used to select representative influencing factors of peak load forecasting error; Error estimation: Based on the preliminary load forecast results of the first adaptive stacking prediction module, the peak load range with large prediction errors can be identified; based on the data from the second adaptive stacking prediction module, the peak load period with prediction errors can be predicted. When retraining the second adaptive stacking prediction module, the original training labels are replaced with the initial training error of the first APM model to estimate the prediction error of multi-energy peak load on the test set. Error correction: The prediction error on the test set is fed back to the first adaptive stacking prediction module to correct the prediction result. The fusion of the intrinsic mode functions of each load type is linearly added. The sum of the preliminary prediction result of the multi-energy load calculated by the first adaptive stacking prediction module and the prediction error output by the second adaptive stacking prediction module is added together.

2. The multi-element, two-stage adaptive stacking prediction method for regional integrated energy systems according to claim 1, characterized in that: Both the first adaptive stacking prediction module and the second adaptive stacking prediction module include three base predictors, one meta predictor, and an optimizer; wherein: the optimizer generates adaptive stacking intermediate parameters by performing a cooperative quantum chaotic search on the load prediction data and load correction data using an atomic chaotic search algorithm, including the following steps: S1-1: Constructing a quantum-based atom search model; S1-2: Determine the initial position of the candidate solution, calculate the initial degree of the candidate solution, randomly generate the number of electron shells, assign the candidate solutions to the electron shells, calculate the binding state, binding energy, and lowest energy level of the electron shell, and generate a random vector; S1-3: The dynamic photon strategy determines whether electrons exhibit photon behavior. Some electrons absorb or undergo changes in their energy state, while others enter chaotic orbits for large-scale searching, thus completing the search for the globally optimal electron in the cooperative atom model.

3. The multi-element two-stage adaptive stacking prediction method for regional integrated energy systems according to claim 2, characterized in that: The process of determining whether an electron exhibits photonic behavior through a dynamic photon strategy, where some electrons absorb photons or change their energy state, and others enter chaotic orbits for large-scale searching, to complete the search for the globally optimal electron in the cooperative atom model, includes the following steps: Step S1-3-1: The energy level of the i-th electron in the k-th electron shell of the dominant atom is greater than the binding state of the dominant atom in the k-th electron shell, i.e.: LO i,k ≥BE L,k The leading electron tends to emit photon energy based on multi-atom cooperation, and the position update equation in the cooperative mode is as follows: LS i,k (λ+1)=LS i,k (l)+a i (2b) i ×LE-γ i ×BS L -s i ×BS F ) / k In the formula: L Si,k (λ) and L Si,k (λ+1) is the λth and (λ+1th)th iteration position of the i-th electron in the k-th shell of the dominant atom, σ i It is a vector randomly generated in the range (0,1), BS L q represents the binding state of the dominant atom. L It is the number of electrons inside the dominant atom, BS F q represents the combination state of all dependent atoms. F It is the number of electrons in the subordinate atom; Otherwise, electrons tend to absorb photons based on multi-atom cooperative absorption, and their cooperative position update equation is as follows: LS i,k (λ+1)=LS i,k (l)+a i (2b) i ×LE L,k -c i ×BS L,k -s i ×BS F,k ) In the formula: d L and d F The total number of electrons in the dominant and subordinate atoms of the kth hypothetical layer; Step S1-3-2: The process of electrons entering chaotic orbits: The chaotic space mapping is established using the following formula: In the formula: s ij (λ) is the j-th decision variable of the i-th candidate solution in the λ-th iteration; s ij,max and s ij,min These are the upper and lower bounds of the j-th decision variable, respectively; assuming S i (λ)=[s i1 (λ),s i2 (λ),...,s ij (λ),...,s im [(λ)] is the position vector of the i-th electron in the chaotic orbit at the λ-th iteration; each dimension s in Si(λ) ij (λ) maps to the chaotic variable c ij (λ)(1≥c ij (λ)≥0); The following iterative logical equation is used for chaotic local search: c ij (λ+1)=μc ij (λ)(1-c) ij (l)) In the formula: μ is the control coefficient. When μ is 4, the chaotic dynamics are characterized by iterative logic equations. Through a sufficient number of iterations, the chaotic orbit of an electron can traverse the entire search space. The chaotic variable c ij (k) Mapping to the original solution space establishes the following solution space mapping: s ij (λ+1)=(s ij,max -s ij,min )c ij (λ+1)+s ij,max Step S1-3-3: Dynamic photon search is performed to meet different energy level optimization requirements by dynamically adjusting PR, including: (1) When the fitness of electron i in the kth iteration is O i (k)≥historical average fitness of electron i O i,ave At this point, the electron under study is far from the lowest energy level, and the photon rate PR can dynamically increase to guide the electron into a chaotic orbit, which can be expressed as: PR=(PR max -PR min ) / 2+r PR (PR max -PR min ) / 2 In the formula PR max and PR min These are the maximum and minimum photon rates, r PR These are random coefficients in [0,1]. (2) When O i (k) <O i,av At this point, the electron under study orbits the electron with the lowest energy level. The electron should then undergo a more refined local search as the iterations proceed. The nonlinear decrease in the photon rate PR can be expressed as: PR max -λ(PR max -PR min ) / λ max In the formula λ max It represents the maximum number of iterations.

4. The multi-element, two-stage adaptive stacking prediction method for regional integrated energy systems according to claim 2, characterized in that: Training process of the base predictor for the first adaptive stacking prediction module and the second adaptive stacking prediction module: The original training data is divided into K folds; three base predictors are trained on the K-1 fold data to predict the remaining one fold data and test data. Each fold of the training set to be predicted corresponds to an independent prediction model; new single-fold training data and a new test set are generated based on the prediction results. By repeatedly performing this iteration, each base predictor can obtain predictions for 5 sets of new training data and 5 sets of new test data. After all base predictors have completed training and prediction, a new training set for a meta predictor is constructed by concatenating the prediction results from the 5-fold training dataset. This meta predictor can be represented as follows: TRS i =[TRP i (1)TRP i (2)TRP i (3)TRP i (4)TRP i (5)] T Where: TRS i TRP represents the new training set generated by the i-th base predictor. i (j) represents the new j-th fold training data predicted by the i-th base predictor; A new test set for a base predictor is obtained by weighting the five sets of predicted test data, which can be represented as: Where: TES i TEP represents the new test set generated by the i-th base predictor. i (j) represents the j-th new test set predicted by the i-th base predictor, W i (j) represents the weight of the j-th new test set of the i-th base prediction; three new training and test datasets containing RF, SVR and LSTM features are transferred to the meta-predictor prediction module for integration.

5. The multi-element, two-stage adaptive stacking prediction method for regional integrated energy systems according to claim 2, characterized in that: The first adaptive stacking prediction module and the second adaptive stacking prediction module perform the following prediction process for the meta-predictor: The three new training datasets obtained from the training of the base predictor are concatenated into a multidimensional training set, and the dimensions of the training set are merged. The original training labels and the merged training set are used in the adaptive boosting method for training; The new test dataset obtained during the training of the base predictor is used to predict future multi-energy load data.

Images