An Energy Demand Forecasting Method Based on Optimizing the PSO-BP Neural Network Model with the ICA Algorithm

Abstract

The present invention relates to the technical field of energy demand prediction, and specifically to an energy demand prediction method based on an ICA algorithm for optimizing a PSO-BP neural network model. First, in order to accurately predict future energy demand, the present invention establishes a future energy impact data generation model based on past data to generate data on the total future energy, energy usage, climate environment, geographical features, population, economic income, energy pollution degree, and energy price. Secondly, the present invention proposes an energy cost model for predicting the cost of the output results of the future energy impact data generation model. Then, in order to more accurately predict future energy demand, the present invention proposes an improved ICA algorithm; the main function of this algorithm is to analyze the influencing factors of energy demand. Finally, the present invention realizes the accurate prediction of future energy demand through an improved PSO-BP network model.

Description

Technical Field

[0001] The present invention relates to the technical field of energy demand forecasting, and specifically to an energy demand forecasting method based on an ICA algorithm for optimizing a PSO-BP neural network model. Background Art

[0002] Energy is one of the necessary substances for the survival and development of human society. Nowadays, various forms of energy are included in human life, such as: petroleum, natural gas, solar energy, wind energy, water energy, and hydrogen energy, etc.

[0003] Nowadays, the use of energy adheres to the concept of sustainable development. Therefore, the task of energy demand forecasting is highly emphasized. A reliable energy demand forecasting method can provide certain data support for the planning of the development strategy of energy enterprises and the formulation of national energy-related policies. Therefore, it is very meaningful for energy enterprises or countries to improve the forecasting accuracy as much as possible.

[0004] In recent years, researchers have combined machine learning and deep learning methods with the energy field and proposed various energy demand forecasting methods. In the existing energy forecasting methods, the laws or trends of various influencing factors changing with time are mainly excavated from historical data, so as to infer future data changes. In this process, in order to achieve more accurate energy demand forecasting, it is necessary to understand the independent contribution degree of each factor. Therefore, in the analysis process of each factor, excellent algorithms in machine learning, such as the ICA algorithm, can be adopted. The ICA algorithm is an algorithm used to separate independent components from mixed signals. In the fields of signal processing and data analysis, ICA is often used to process mixed signals, where different components are mixed together in an independent manner. Although ICA is mainly used in the field of signal processing, in energy demand analysis, different influencing factors of energy demand can be considered as mixed signals, and by applying the ICA algorithm, an attempt is made to separate independent components from them. This can help identify and understand the contributions of different influencing factors to the overall energy demand. The application of the ICA algorithm in the energy forecasting field is mainly divided into the following two forms: one is to establish a forecasting model only using historical energy demand data, that is, single-factor time series forecasting; the other is to establish a forecasting model based on multi-factor influence using historical factor data affecting energy demand, such as factors like gross domestic product, population quantity, and import and export volume, etc. However, these two methods only focus on the demand forecasting of a single energy source and do not comprehensively consider the demands of various types of energy.

[0005] To solve the above problems, the present invention proposes an energy demand forecasting method based on an ICA algorithm for optimizing a PSO-BP neural network model. Summary of the Invention

[0006] An energy demand prediction method based on the ICA algorithm to optimize the PSO - BP neural network model is used to solve the above - mentioned technical problems. First, in order to accurately predict future energy demand, the present invention establishes a future energy impact data generation model based on past data to generate the total energy volume, energy usage, climate environment, geographical features, population, economic income, energy pollution degree, and energy price in the future. Secondly, the present invention proposes an energy cost model to predict the cost of the output results of the future energy impact data generation model. Then, in order to more accurately predict future energy demand, the present invention proposes an improved ICA algorithm; the main function of this algorithm is to analyze the influencing factors of energy demand. Finally, the present invention realizes the accurate prediction of future energy demand through an improved PSO - BP network model.

[0007] An energy demand prediction method based on the ICA algorithm to optimize the PSO - BP neural network model includes:

[0008] Collect historical energy impact data and historical energy cost data in the region in the past N years; where N is a natural number greater than 1; among them, the energy impact data includes: total energy volume, energy usage, climate environment, geographical features, population, economic income, energy pollution degree, and energy price.

[0009] Furthermore, pre - process the historical energy impact data to obtain standard energy impact data and construct a standard energy impact data set. Among them, the specific process of the pre - processing includes: clean the energy impact data to obtain the first energy impact data; where the data cleaning processes the missing values, outliers, and error values in the energy impact data; perform time - series processing on the first energy impact data to obtain the second energy impact data; where the time - series processing includes: decomposition by date, seasonal adjustment, and trend analysis; the second energy impact data is processed by standardization and normalization to obtain the third energy impact data; perform one - hot encoding on the second energy impact data to obtain the standard energy impact data.

[0010] Furthermore, establish an energy cost impact correlation matrix between the standard energy impact data set and the energy cost; where the energy cost impact correlation matrix is expressed as:

[0011]

[0012] Among them, EPICM represents the energy cost impact correlation matrix; I 1K represents the Kth energy cost impact index in the first piece of historical energy impact data; EC 1M represents the Mth type of energy cost in the first piece of historical energy impact data; INM Denoted as the K-th energy cost impact indicator in the N-th historical energy impact data; EC NM Denoted as the M-th type of energy cost in the N-th historical energy impact data.

[0013] Furthermore, calculate the differences in historical energy impact data according to the energy cost impact correlation matrix, and construct a historical energy impact difference matrix where HERDDF is denoted as the historical energy impact difference matrix; Denoted as the difference between the K-th energy cost impact indicator in the N-th data of the energy cost impact correlation matrix and the K-th energy cost impact indicator in the (N - 1)-th data; The calculation formula of is:

[0014] Furthermore, initialize the energy impact change vector W = (w1,..., w K ); where w K Denoted as the change value of the K-th energy cost impact indicator,

[0015] Furthermore, establish a future energy impact data generation model according to the energy cost impact correlation matrix and the initialized energy cost impact correlation matrix, and obtain future energy impact data.

[0016] Furthermore, input the future energy impact data into the energy cost model to obtain the future energy prediction cost; where the specific implementation of the energy cost model includes:

[0017] Normalize the historical energy cost data for N years to obtain normalized historical energy cost data;

[0018] Cluster according to the normalized historical energy cost data to obtain C clustering sets; where C is a natural number greater than 1;

[0019] Construct C relevant energy data difference matrices for the C clustering sets; where the relevant energy data difference matrix is denoted as: where EDGM i Denoted as the relevant energy data difference matrix of the i-th clustering set; DV QK Denoted as the difference value of the Q-th data in the i-th clustering set with respect to the K-th factor;

[0020] Establish an energy cost impact factor transfer matrix according to the clustering sets; where the energy cost impact factor transfer matrix is denoted as: where EPFITM iThe transfer matrix of energy cost influencing factors represented as the i-th clustering set; p 1K It represents the transfer probability value between the first energy influencing factor and the K-th energy influencing factor in the i-th clustering set;

[0021] The constraint condition of the transfer matrix of energy cost influencing factors is:

[0022] Update the influencing factor weight vector according to the transfer matrix of energy cost influencing factors; among them, the update formula of the influencing factor weight vector is:

[0023]

[0024] Among them, μ i (t) represents the weight vector of cost influencing factors of the i-th clustering set at time t; ε(t) represents the correction term at time t;

[0025] Establish an energy cost prediction function; among them, the formula of the energy cost prediction function is:

[0026]

[0027] Among them, EP() i represents the energy cost prediction function of the i-th clustering set; BEC m represents the basic cost of the M-th type of energy; φ() represents the non-linear part of the cost prediction function; B represents the bias of the energy cost prediction function;

[0028] Among them, the update formula of the bias is: λ represents the correction parameter;

[0029] Minimize the energy cost prediction function; among them, the minimization process is:

[0030]

[0031] Among them, T represents the maximum number of iterative training rounds; EP t m represents the prediction result of the energy cost prediction function for the m-th type of energy in the t-th round;

[0032] Furthermore, calculate the loss value between the future energy prediction cost and each piece of the historical energy cost data, and optimize the energy cost model;

[0033] Among them, the calculation formula of the loss value is:

[0034]

[0035] Wherein, N represents the total number of historical energy impact data; M represents the total number of energy categories; represents the predicted cost of the m-th type of energy; represents the actual cost of the m-th type of energy.

[0036] Further, clustering the future energy impact data with the clustering set; obtaining a clustering result;

[0037] Inputting the future energy impact data into the energy cost prediction function of the corresponding clustering set to obtain the future energy prediction cost.

[0038] Further, combining the standard energy impact data, the historical energy cost, and the historical energy development policy feature vector to obtain an energy demand feature weight matrix;

[0039] Wherein, the specific process of obtaining the energy development policy features includes:

[0040] Obtaining relevant energy development policy documents issued in the region;

[0041] Extracting the energy-related content in the relevant energy development policy documents;

[0042] Wherein, the energy-related content includes: energy status information, energy planning information, environmental impact information, economic fluctuation information, and energy support information;

[0043] Using feature engineering methods for the energy-related content to obtain the energy development policy feature vector.

[0044] The energy demand feature weight matrix adopts an energy demand feature engineering analysis model;

[0045] Wherein, the specific implementation process of the energy demand feature engineering analysis model includes: constructing an energy demand feature weight matrix; wherein, the energy demand feature weight matrix is expressed as: MDMM=(HD1;...; HD i ;...; HD N ) Τ ; wherein, HD i represents the vector composed of the standard energy impact data, the historical energy cost data, and the historical energy development policy feature vector in the i-th year;

[0046] Normalizing the energy demand feature weight matrix to obtain a normalized energy demand feature weight matrix;

[0047] Establishing an energy demand feature analysis function; wherein, the expression of the energy demand feature analysis function is:

[0048]

[0049] Among them, EDCA(t) represents the energy demand characteristic analysis function at time t; X(t) represents the observation result of energy demand characteristics; It is expressed as the time characteristic analysis function;

[0050] The calculation formula of the said X(t) is:

[0051] X(t) = ω * A * MDMM;

[0052] Among them, it represents the energy demand characteristic weight matrix; A represents the mixing matrix;

[0053] The formula of the said time characteristic analysis function is:

[0054]

[0055] Among them, β0, β1,..., β n It represents the parameters of the said time characteristic analysis function; ▽δ represents the error term.

[0056] Furthermore, input the future energy impact data, the future energy prediction cost, and the future energy development policy feature vector into the energy demand prediction model;

[0057] Among them, the specific implementation steps of the said energy demand prediction model include:

[0058] Step 1: Determine the energy demand prediction model structure according to the standard energy impact data set, the historical energy cost data, and the historical energy development policy feature vector;

[0059] Step 2: Obtain the energy demand characteristic weight matrix as the initialization weight parameter of the energy demand prediction model;

[0060] Step 3: Optimize the particle swarm algorithm according to the energy demand, and initialize the velocity, position, individual extreme value, and global extreme value of the particles;

[0061] Among them, the specific process of the said energy demand optimization particle swarm algorithm includes:

[0062] Set a particle swarm; among them, is a natural number greater than 1;

[0063] Initialize the position set and velocity set of the particles; among them, the position set is expressed as L i =(l i1 , l i2 ,..., l iD ); l iD represents the position of the D-th particle in the i-th particle swarm; the velocity set is expressed as V i =(vi1 , v i2 ,..., v iD ); v iD It represents the velocity of the D-th particle in the i-th particle swarm;

[0064] Update the position set and the velocity set of the particles; where the update formula is:

[0065]

[0066] Where It represents the velocity of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration; It represents the velocity of the D-th particle in the i-th particle swarm at the k-th iteration; It represents the position of the D-th particle in the i-th particle swarm at the k-th iteration; c1 and c2 represent learning factors that are non-negative constants; p iD It represents the local optimal value of the D-th particle in the i-th particle swarm; p gD It represents the global optimal value of the D-th particle; r1 k , It represents a random number in [0, 1] at the k-th iteration; θ(k) represents the inertia weight coefficient; It represents the position of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration;

[0067] Where the calculation formula of the inertia weight coefficient is as follows:

[0068]

[0069] Where t represents the number of iterations; θ max It represents the maximum value of the inertia weight coefficient; θ min It represents the minimum value of the inertia weight coefficient; t max It represents the maximum number of iterations; d represents the initial inertia weight after the initial search.

[0070] Step 4: Select a fitness function, evaluate the fitness value of each particle, and obtain the initialized particle fitness value set;

[0071] Step 5: Evaluate each element of the initialized particle fitness value set; if the current fitness value is better than the local optimal solution,

[0072] Step 6: Update the local optimal solution; if the current fitness value is better than the global optimal solution, update the global optimal solution;

[0073] Step 7: Recalculate the velocity and speed of the particles and perform a mutation operation;

[0074] Step 8: Determine whether the number of iterations is less than the preset value. If so, return to Step 4;

[0075] Step 9: Allocate the obtained optimal value to the model for training and learning.

[0076] Output the future energy demand prediction result.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0078] 1. The present invention proposes a future energy impact data generation model, which can generate future energy impact data based on the collected historical data. Among them, during the model construction process, a correlation matrix is established through historical total energy, energy consumption, climate environment, geographical features, population quantity, economic income, energy pollution degree, and energy price, and data is generated by combining the energy impact change vectors of multiple factors. The future energy impact data generation model realizes the generation of future data based on real data, has a certain rationality, provides scientific data for subsequent energy demand prediction, and improves the prediction accuracy.

[0079] 2. The present invention proposes an energy cost model, which is used to further predict the output results of the future energy impact data generation model. This model mainly establishes a difference matrix of influencing factors for each category through clustering, and this matrix can reflect the cost changes between various factors, improving the accuracy of energy cost prediction within the same set. In addition, it is one of the important influencing factors indispensable for the prediction cost of energy demand. Therefore, accurate cost prediction is beneficial to improving the accuracy of energy demand prediction.

[0080] 3. The present invention proposes an energy demand feature engineering analysis model, which improves on the basic ICA algorithm. This model adds a time feature analysis function to deepen the analysis of influencing factors. In addition, when this model was designed, considering that energy demand is affected by relevant energy policies in the actual scenario, necessary features related to energy development in the policies are extracted for analysis to obtain the weights of influencing factors, thereby improving the accuracy of the final energy demand prediction.

[0081] 4. The present invention proposes an energy demand prediction model, which is used to predict future energy demand. This model optimizes the parameters of the BP neural network based on the energy demand optimized particle swarm algorithm. The energy demand optimized particle swarm algorithm is an improvement on the particle swarm algorithm, mainly by optimizing the mutation strategy of the inertia weight, ensuring the global search ability in the early stage of the algorithm and the rapid convergence to the optimal solution in the later stage. An adaptive mutation algorithm is introduced during the search process to prevent particles from falling into local optima. This optimization algorithm can well optimize network parameters and improve the accuracy of energy demand prediction. Description of the Drawings

[0082] Figure 1 This is the chart of the proportion of energy use in City A in 2023 provided by the embodiments of the present invention;

[0083] Figure 2 This is the flowchart of an energy demand prediction method based on an ICA algorithm-optimized PSO-BP neural network model provided by the embodiments of the present invention;

[0084] Figure 3 This is the flowchart of the energy demand prediction model provided by the embodiments of the present invention. Detailed implementation manners

[0085] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0086] Energy is an important material basis for people's survival and the promotion of social civilization development. With the continuous growth of population and economy, energy demand is also increasing continuously. Therefore, the research on energy demand has important theoretical significance and practical significance. In the prior art, most of the predictions are made for the demand of single types of energy, such as natural gas or electric energy. However, these methods cannot well reflect the diversity of energy demand and are not conducive to the rationing supply of energy. For such problems, the present invention proposes an energy demand prediction method based on an ICA algorithm-optimized PSO-BP neural network model, which can well reflect the demand predictions of various types of energy and provide reasonable data support for energy supply and distribution. To further illustrate the feasibility of the present invention in actual situations, the following embodiments will be described in detail.

[0087] Embodiment 1

[0088] In the embodiments of the present application, by analyzing the energy situation in City A in the past 10 years and using the method of the present invention to verify the total energy demand in 2023; the main energy supply and demand in City A include: nuclear energy, hydropower, petroleum, renewable energy, natural gas and coal. In Figure 1 shows the proportion of energy use in City A in 2023. To describe the prediction process of energy demand in detail, it will be achieved through the following steps.

[0089] Refer to Figure 2, The specific implementation process of an energy demand prediction method based on the ICA algorithm to optimize the PSO-BP neural network model includes: S10. Obtain the historical energy impact data and historical energy cost data of City A in the past 10 years; S20. Preprocess the historical energy impact data to construct a historical energy information dataset; S30. Use the future energy impact data generation model to obtain the future relevant energy data vector; S40. Use the energy cost model to obtain the future energy prediction cost; S50. Calculate the energy demand characteristic weight matrix; S60. Use the energy demand prediction model to make a prediction; S70. Output the future energy demand prediction result. The following will provide a detailed process description of the above-mentioned steps.

[0090] In the above S10, the historical energy impact data and historical energy cost data of City A in the past 10 years are collected; among them, the energy impact data includes: total energy, energy usage, climate environment, geographical features, population, economic income, energy pollution degree, and energy price.

[0091] In order to accurately predict the energy demand of City A, a series of histories related to energy are collected in the embodiments of the present application. Through the analysis of historical data, the future energy demand can be more accurately inferred. Moreover, the above-listed related data types have indirect or direct impacts on the fluctuations of energy demand. Therefore, energy demand prediction under multi-factor conditions not only meets the needs of the real scenario but also provides sufficient data support for the present invention.

[0092] Furthermore, in order to illustrate the energy usage situation of City A in recent years, in the embodiments of the present application, the energy usage situation of City A in the past 5 years is obtained. Refer to the content in Table 1.

[0093] Table 1. Energy usage proportion of City A in the past 5 years

[0094] Year Nuclear energy (%) Hydropower (%) Oil (%) Renewable energy (%) Natural gas (%) Coal (%) 2019 0.17 6.52 7.22 1.61 3.06 81.42 2020 0.23 6.08 7.16 2.14 3.74 80.65 2021 0.45 6.00 7.63 2.35 4.68 78.89 2022 0.80 8.43 10.41 2.78 5.24 72.34 2023 1.00 8.00 17.00 3.00 6.00 65.00

[0095] Furthermore, the above-mentioned S20 step is used to preprocess the collected historical energy impact data; the data preprocessing process includes:

[0096] Data cleaning is performed on the energy impact data to obtain the first energy impact data; among them, the data cleaning processes the missing values, outliers, and error values in the energy impact data; time series processing is performed on the first energy impact data to obtain the second energy impact data; among them, the time series processing includes: decomposition by date, seasonal adjustment, and trend analysis; the second energy impact data is processed by standardization and normalization to obtain the third energy impact data; one-hot encoding is performed on the second energy impact data to obtain the standard energy impact data.

[0097] During the process of collecting relevant data, due to the large amount of data, there are errors in the data, which is not conducive to subsequent analysis. Therefore, in the embodiments of the present application, the data is preprocessed to improve the data resolvability. Among them, the preprocessing process includes: data cleaning, time series processing, and standardization and normalization processing. Data cleaning can remove errors, outliers, and missing values in the dataset, thereby improving the quality of the data. This helps to ensure that the model is built on accurate and reliable data, avoiding inaccurate prediction results caused by data quality problems; energy demand is usually affected by seasonal and cyclical factors, such as weather seasons, industrial production cycles, etc. Time series analysis can effectively capture and simulate these trends, thereby improving the prediction accuracy of future energy demand; standardization and normalization can eliminate the influence caused by different scales between different features. In energy demand prediction, multiple features may be involved, such as temperature, population, economic indicators, etc., and their value ranges may be different. Standardization and normalization can scale the values of these features to a similar scale, making the model more stable and accurate.

[0098] Further, according to the standard energy impact data, a standard energy impact dataset is constructed.

[0099] Further, in the above step S30, the standard energy impact dataset is used to generate relevant energy data for 2023; among them, the generation method of the relevant data for 2023 is obtained by using a future energy impact data generation model;

[0100] The specific implementation process of the future energy impact data generation model includes:

[0101] An energy cost impact correlation matrix is established according to the standard energy impact dataset and the historical energy cost data of City A; among them, the energy cost impact correlation matrix is expressed as:

[0102]

[0103] Among them, EPICM represents the energy cost impact correlation matrix; I 1K represents the Kth energy cost impact index in the first historical energy impact data; EC 1M represents the Mth type of energy cost in the first historical energy impact data; I NM represents the Kth energy cost impact index in the Nth historical energy impact data; EC NM represents the Mth type of energy cost in the Nth historical energy impact data.

[0104] In the embodiments of the present application, the energy cost impact correlation matrix is established to better analyze the relationships between various influencing factors and various types of energy. This matrix is composed of the mapping relationship between the energy impact data and the energy cost in the historical data. Through this matrix, the conditional factors for the generation of energy costs can be further analyzed.

[0105] Further, calculate the differences in historical energy impact data according to the energy cost impact correlation matrix, and construct a historical energy impact difference matrix. Among them, HERDDF represents the historical energy impact difference matrix. represents the difference between the Kth energy cost impact index in the Nth data in the energy cost impact correlation matrix and the Kth energy cost impact index in the (N - 1)th data. The calculation formula of is:

[0106] Further, initialize the energy impact change vector W=(w1,...,w K ) according to the historical energy impact difference matrix; where, w K represents the change value of the Kth energy cost impact index.

[0107] Further, establish a future energy impact data generation model according to the energy cost impact correlation matrix and the initialized energy cost impact correlation matrix, and obtain future energy impact data.

[0108] Further, as in step S40 above, input the future energy impact data into the energy cost model to obtain the future energy prediction cost; where, the specific implementation of the energy cost model includes:

[0109] Normalize the historical energy cost data for N years to obtain normalized historical energy cost data.

[0110] Cluster according to the normalized historical energy cost data to obtain C clustering sets; where, C is a natural number greater than 1.

[0111] Construct C relevant energy data difference matrices for the C clustering sets; where, the relevant energy data difference matrix is expressed as: Among them, EDGM i represents the relevant energy data difference matrix of the ith clustering set; DV QK represents the difference value of the Qth data in the ith clustering set regarding the Kth factor.

[0112] Establish an energy cost impact factor transfer matrix according to the clustering sets; where, the energy cost impact factor transfer matrix is expressed as: Among them, EPFITM i represents the transfer matrix of energy cost influencing factors for the i-th clustering set; p 1K represents the transfer probability value between the first energy influencing factor and the K-th energy influencing factor in the i-th clustering set; the constraint conditions of the transfer matrix of energy cost influencing factors are:

[0113] Cluster the data collected in City A and construct the transfer matrix of energy cost influencing factors according to the energy cost influencing factors as Among them, TEC represents the total energy; EU represents the energy consumption; CE represents the climate environment; GC represents the geographical features; PS represents the population; EI represents the economic income; EPL represents the energy pollution degree; EP represents the energy price.

[0114] Update the influencing factor weight vector according to the transfer matrix of energy cost influencing factors; among them, the update formula of the influencing factor weight vector is:

[0115]

[0116] Among them, μ i (t) represents the weight vector of cost influencing factors for the i-th clustering set at time t; ε(t) represents the correction term at time t;

[0117] Establish an energy cost prediction function; among them, the formula of the energy cost prediction function is:

[0118]

[0119] Among them, EP() i represents the energy cost prediction function for the i-th clustering set; BEC m represents the basic cost of the M-th type of energy; φ() represents the non-linear part of the cost prediction function; B represents the bias of the energy cost prediction function;

[0120] Among them, the update formula of the bias is: λ represents the correction parameter;

[0121] Minimize the energy cost prediction function; among them, the minimization process is:

[0122]

[0123] Among them, T represents the maximum number of iterative training rounds; EP t m represents the prediction result of the energy cost prediction function for the m-th type of energy in the t-th round;

[0124] Further, calculate the loss value between the future energy prediction cost and each type of historical energy cost data, and optimize the energy cost model;

[0125] Among them, the calculation formula of the loss value is:

[0126]

[0127] Among them, N represents the total number of historical energy impact data; M represents the total number of energy categories; represents the predicted cost of the m-th type of energy; represents the actual cost of the m-th type of energy.

[0128] In the embodiment of the present application, loss calculation is performed on the established energy cost model during the training process. This loss calculation formula can measure the difference between the model prediction result and the actual observed value. Using a suitable loss function to train and evaluate the model brings many benefits; for example, guiding the optimization of the energy cost model and guiding the convergence of the energy cost model.

[0129] In the embodiment of the present application, an energy cost model is established. The main function of this model is to perform cost prediction by combining the generated energy impact data. During the analysis of energy demand, it is found that energy cost is one of the indispensable important factors. Therefore, predicting the future energy cost is beneficial to improving the accuracy of predicting the future energy demand.

[0130] Further, cluster the future energy impact data with the clustering set; obtain the clustering result;

[0131] Input the future energy impact data into the energy cost prediction function of the corresponding clustering set to obtain the future energy prediction cost.

[0132] Further, as in S50 above, combine the standard energy impact data, the historical energy cost, and the historical energy development policy feature vector to obtain an energy demand feature weight matrix;

[0133] Among them, the specific process of obtaining the energy development policy features includes:

[0134] Obtain relevant energy development policy documents issued in the region;

[0135] Extract the energy-related content in the relevant energy development policy documents;

[0136] Among them, the energy-related content includes: energy status information, energy planning information, environmental impact information, economic fluctuation information, and energy support information;

[0137] Adopt feature engineering methods for the energy-related content to obtain the energy development policy feature vector.

[0138] There is a close relationship between the supply and demand of energy and relevant energy policies, which guide and manage the development, utilization, distribution, and protection of energy resources by countries, regions, or organizations. Therefore, in order to conform to the use of energy in the real scenario, the present invention combines the energy demand with relevant policies and extracts the necessary energy-related content from the relevant policies; among them, it includes: energy status information, energy planning information, environmental impact information, economic fluctuation information, and energy support information; among these, feature engineering technology is adopted for feature vectorization. Through this method, energy demand prediction can be associated with corresponding policies, conforming to the real scenario. Further, the energy demand feature weight matrix adopts an energy demand feature engineering analysis model;

[0139] Among them, the specific implementation process of the energy demand feature engineering analysis model includes: constructing an energy demand feature weight matrix; among them, the energy demand feature weight matrix is expressed as: MDMM = (HD1;...; HD i ;...; HD N ) Τ ; among them, HD i represents a vector composed of the standard energy impact data, the historical energy cost data, and the historical energy development policy feature vector in the i-th year;

[0140] Standardize the energy demand feature weight matrix to obtain a standardized energy demand feature weight matrix;

[0141] Establish an energy demand feature analysis function; among them, the expression of the energy demand feature analysis function is:

[0142]

[0143] Among them, EDCA(t) represents the energy demand feature analysis function at time t; X(t) represents the energy demand feature observation result; represents the time feature analysis function;

[0144] The calculation formula of the said X(t) is:

[0145] X(t) = ω * A * MDMM;

[0146] Among them, ω represents the energy demand feature weight matrix;

[0147] The formula of the time feature analysis function is:

[0148]

[0149] where, β0, β1, ..., β n represent the parameters of the time feature analysis function; ▽δ represents the error term.

[0150] In the embodiments of the present application, an energy demand feature engineering analysis model proposed by the present invention is used to comprehensively analyze energy influencing factors, and different weight mappings are assigned to different factors. This model is an improved algorithm based on ICA; where ICA is the English abbreviation for Independent Component Analysis, and the Chinese name is independent component analysis; this algorithm is a mathematical and statistical method used to decompose the mixed signals of multiple random variables into independent non-Gaussian signal sources. It is often used in the fields of signal processing, data analysis, and machine learning. In the present invention, by adding a high-order regression function, the independent components and their dynamic properties in time series data can be analyzed and understood more comprehensively and deeply.

[0151] Further, refer to Figure 3 , as in the above step S60, input the future energy impact data, the future energy prediction cost, and the future energy development policy feature vector into the energy demand prediction model;

[0152] In the embodiments of the present application, the energy demand prediction model proposed by the present invention is used for energy demand prediction. This model combines an improved particle swarm algorithm with a BP neural network for prediction. It can be seen from the verification of the energy demand results in 2023 that the error with the actual quantity in Table 2 below is small, and the effectiveness of the energy demand prediction model can be obtained.

[0153] Among them, the specific implementation steps of the energy demand prediction model include:

[0154] Step 1: Determine the energy demand prediction model structure according to the standard energy impact data set, the historical energy cost data, and the historical energy development policy feature vector;

[0155] Step 2: Obtain the energy demand feature weight matrix as the initialization weight parameter of the energy demand prediction model;

[0156] Step 3: Optimize the particle swarm algorithm according to the energy demand, and initialize the velocity, position, individual extreme value, and global extreme value of the particles;

[0157] Among them, the specific process of the energy demand optimized particle swarm algorithm includes:

[0158] Set E particle swarms; where E is a natural number greater than 1;

[0159] Initialize the position set and velocity set of the particles; where the position set is represented as L i =(li1 , l i2 ,..., l iD ); l iD denotes the position of the D-th particle in the i-th particle swarm; the velocity set is denoted as V i = (v i1 , v i2 ,..., v iD ); v iD denotes the velocity of the D-th particle in the i-th particle swarm;

[0160] Update the position set and the velocity set of the particles; where the update formula is:

[0161]

[0162] where, denotes the velocity of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration; denotes the velocity of the D-th particle in the i-th particle swarm at the k-th iteration; denotes the position of the D-th particle in the i-th particle swarm at the k-th iteration; c1 and c2 denote learning factors of non-negative constants; p iD denotes the local optimal value of the D-th particle in the i-th particle swarm; p gD denotes the global optimal value of the D-th particle; r1 k , denotes a random number in [0, 1] at the k-th iteration; θ(k) denotes the inertia weight coefficient; denotes the position of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration;

[0163] where, the calculation formula of the inertia weight coefficient is as follows:

[0164]

[0165] where, t denotes the number of iterations; θ max denotes the maximum value of the inertia weight coefficient; θ min denotes the minimum value of the inertia weight coefficient; t max denotes the maximum number of iterations; d denotes the initial inertia weight after the initial search.

[0166] Step 4: Select a fitness function, evaluate the fitness value of each particle, and obtain an initialized particle fitness value set;

[0167] Step 5: Evaluate each element of the initialized particle fitness value set; if the current fitness value is better than the local optimal solution,

[0168] Step 6: Update the local optimal solution; if the current fitness value is better than the global optimal solution, update the global optimal solution;

[0169] Step 7: Recalculate the velocity and position of the particle and perform the mutation operation;

[0170] Step 8: Determine whether the number of iterations is less than the preset value, and return to Step 4;

[0171] Step 9: Assign the obtained optimal value to the model for training and learning;

[0172] In the energy demand prediction model, the mutation strategy of the inertia weight is optimized by the improved particle swarm algorithm, which ensures the global search ability in the early stage of the algorithm and the rapid convergence to the optimal solution in the later stage. The adaptive mutation algorithm is introduced in the search process to prevent the particles from falling into the local optimum.

[0173] As described in S70 above, output the future energy demand prediction result.

[0174] In the implementation of this application, the energy demand of City A in 2023 was verified. To illustrate that the prediction result of the energy demand prediction model has good accuracy, the following Table 2 shows through data that the model made an accurate prediction;

[0175] Table 2. Energy usage in City A for the past 5 years

[0176] Year Total energy consumption (10,000 tons of standard coal) 2019 68732 2020 74569 2021 79432 2022 80534 2023 82657

[0177] According to the output result in Step S70, the energy consumption in 2023 is approximately 819.473 million tons of standard coal, and the error from the actual result in the table is approximately 7.097 million tons of standard coal.

[0178] In the first embodiment, an energy demand prediction method based on the ICA algorithm to optimize the PSO-BP neural network model proposed by the present invention is used to verify the energy demand of City A in 2023. This method mainly conducts demand prediction through the following four core models; First, the future energy impact data generation model is used to speculate on the future influencing factor data and generate future data based on the existing data foundation; Second, the energy cost model establishes a connection according to the existing data and estimates the cost of the generated future data. Among them, energy cost is one of the major factors for energy prediction; Then, the energy demand feature engineering analysis model is used to comprehensively analyze the energy factors and establish an energy demand feature weight matrix for initializing the weights of the demand model; Finally, the energy demand prediction model is used for energy demand prediction, and this model combines the improved particle swarm algorithm with the BP neural network to obtain accurate prediction results.

[0179] Embodiment 2

[0180] In the above-mentioned first embodiment, the solution of the present invention is adopted for effect verification. It can be seen from the final results that the present invention is feasible. According to the operations in the first embodiment, the energy demand of City A in 2024 will be predicted in the second embodiment, and the process is as follows:

[0181] Collect the historical energy impact data and historical energy cost data of City A in the past 10 years; among them, the 10-year historical data includes from 2014 to 2023.

[0182] The energy impact data includes: total energy, energy usage, climate environment, geographical features, population quantity, economic income, energy pollution degree, and energy price.

[0183] Furthermore, preprocess the data to obtain standard energy impact data and construct a data set;

[0184] Furthermore, use the future energy impact data generation model to generate the energy impact data in 2024;

[0185] Establish an energy cost impact correlation matrix according to the standard energy impact data set and the historical energy cost data of City A; among them, the energy cost impact correlation matrix is expressed as:

[0186]

[0187] Among them, EPICM represents the energy cost impact correlation matrix; I 1K represents the Kth energy cost impact index in the first piece of historical energy impact data; EC 1M represents the Mth type of energy cost in the first piece of historical energy impact data; I NM represents the Kth energy cost impact index in the Nth piece of historical energy impact data; EC NM represents the Mth type of energy cost in the Nth piece of historical energy impact data.

[0188] Furthermore, calculate the differences in historical energy impact data according to the energy cost impact correlation matrix and construct a historical energy impact difference matrix Among them, HERDDF represents the historical energy impact difference matrix; represents the difference between the Kth energy cost impact index in the Nth data in the energy cost impact correlation matrix and the Kth energy cost impact index in the (N - 1)th data; The calculation formula of is:

[0189] Furthermore, initialize the energy impact change vector W=(w1,...,w K ); among them, wK Denoted as the change value of the Kth energy cost impact indicator,

[0190] Furthermore, according to the energy cost impact correlation matrix and the initialized energy cost impact correlation matrix, a future energy impact data generation model is established, and the energy impact data vector for 2024 is obtained.

[0191] Furthermore, the energy impact data vector for 2024 is input into the energy cost model to obtain the energy cost for 2024;

[0192] Furthermore, the standard energy impact data, the historical energy cost, and the historical energy development policy feature vector are combined to obtain an energy demand feature weight matrix;

[0193] Among them, the specific process of obtaining the energy development policy features includes:

[0194] Obtain relevant energy development policy documents issued in the region;

[0195] Extract the energy-related content in the relevant energy development policy documents;

[0196] Among them, the energy-related content includes: energy status information, energy planning information, environmental impact information, economic fluctuation information, and energy support information;

[0197] Adopt feature engineering methods for the energy-related content to obtain the energy development policy feature vector.

[0198] Furthermore, input the energy impact data for 2024, the predicted energy cost for 2024, and the future energy development policy feature vector into the energy demand prediction model;

[0199] Among them, the specific implementation steps of the energy demand prediction model include:

[0200] Step 1: Determine the energy demand prediction model structure according to the standard energy impact data set, the historical energy cost data, and the historical energy development policy feature vector;

[0201] Step 2: Obtain the energy demand feature weight matrix as the initialized weight parameter of the energy demand prediction model;

[0202] Step 3: Initialize the velocity, position, individual extreme value, and global extreme value of the particles according to the energy demand optimization particle swarm algorithm;

[0203] Among them, the specific process of the energy demand optimization particle swarm algorithm includes:

[0204] Set up E particle swarms; where E is a natural number greater than 1;

[0205] Initialize the position set and velocity set of the particles; where the position set is denoted as L i =(l i1 ,l i2 ,...,l iD ); l iD represents the position of the D-th particle in the i-th particle swarm; the velocity set is denoted as V i =(v i1 ,v i2 ,...,v iD ); v iD represents the velocity of the D-th particle in the i-th particle swarm;

[0206] Update the position set and the velocity set of the particles; where the update formula is:

[0207]

[0208] where, represents the velocity of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration; represents the velocity of the D-th particle in the i-th particle swarm at the k-th iteration; represents the position of the D-th particle in the i-th particle swarm at the k-th iteration; c1 and c2 represent learning factors of non-negative constants; p iD represents the local optimal value of the D-th particle in the i-th particle swarm; p gD represents the global optimal value of the D-th particle; r1 k , represents a random number in [0, 1] at the k-th iteration; θ(k) represents the inertia weight coefficient; represents the position of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration;

[0209] where, the calculation formula of the inertia weight coefficient is as follows:

[0210]

[0211] where, t represents the number of iterations; θ max represents the maximum value of the inertia weight coefficient; θ min represents the minimum value of the inertia weight coefficient; t max represents the maximum number of iterations; d represents the initial inertia weight after the initial search.

[0212] Step 4: Select a fitness function, evaluate the fitness value of each particle, and obtain the initialized particle fitness value set;

[0213] Step Five: Evaluate each element in the initialized particle fitness value set; if the current fitness value is better than the local optimal solution,

[0214] Step Six: Update the local optimal solution; if the current fitness value is better than the global optimal solution, update the global optimal solution;

[0215] Step Seven: Recalculate the velocity and position of the particle, and perform a mutation operation;

[0216] Step Eight: Determine whether the number of iterations is less than the preset value, and return to Step Four;

[0217] Step Nine: Assign the obtained optimal value to the model for training and learning;

[0218] Output the prediction result of the energy demand in 2024.

[0219] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An energy demand forecasting method based on optimizing the PSO-BP neural network model with the ICA algorithm, characterized in that, Including: Collecting historical energy impact data and historical energy cost data for N years in the region where it is located; N is a natural number greater than 1; Preprocessing the historical energy impact data to obtain standard energy impact data and constructing a standard energy impact data set; Establishing an energy cost impact correlation matrix between the standard energy impact data set and the historical energy cost data; Establish a historical energy impact difference matrix ; among them, represents the difference value in the (N - 1)-th row and K-th column of the historical energy impact difference matrix; The calculation formula of is: represents the K-th energy cost impact index in the N-th historical energy impact data; Initialize the energy impact change vector ; where represents the change value of the K-th energy cost impact index ; Constructing a future energy impact data generation model based on the energy cost impact correlation matrix and the energy impact change vector to obtain future energy impact data; Training an energy cost model using the standard energy impact data set, calculating a loss value, and optimizing the model; Inputting the future energy impact data into the energy cost model to obtain a future energy prediction cost; Combining the standard energy impact data, the historical energy cost, and the historical energy development policy feature vector to obtain an energy demand feature weight matrix; The energy demand feature weight matrix adopts an energy demand feature engineering analysis model; Among them, the specific implementation process of the energy demand feature engineering analysis model includes: constructing an energy demand feature weight matrix; among them, the energy demand feature weight matrix is expressed as: ; among them, is expressed as a vector composed of the standard energy impact data, the historical energy cost data, and the historical energy development policy feature vector in the i-th year; Standardizing the energy demand feature weight matrix to obtain a standardized energy demand feature weight matrix; Establishing an energy demand feature analysis function; wherein, the expression of the energy demand feature analysis function is: ； Among them, represents the energy demand characteristic analysis function at time t; represents the observed result of energy demand characteristics; represents the time characteristic analysis function; The said The calculation formula is as follows: ； Among them, is expressed as the energy demand characteristic weight matrix; A is expressed as the mixing matrix; The formula of the time feature analysis function is: ； Among them, is expressed as a parameter of the time feature analysis function; is expressed as an error term; Inputting the future energy impact data, the future energy prediction cost, and the future energy development policy feature vector into an energy demand prediction model; The specific implementation steps of the energy demand prediction model include: Step 1: Determining the energy demand prediction model structure according to the standard energy impact data set, the historical energy cost data, and the historical energy development policy feature vector; Step 2: Obtaining the energy demand feature weight matrix as the initial weight parameter of the energy demand prediction model; Step 3: Initializing the velocity, position, individual extreme value, and global extreme value of the particles according to the energy demand optimization particle swarm algorithm; Step 4: Selecting a fitness function to evaluate the fitness value of each particle to obtain an initialized particle fitness value set; Step 5: Evaluating each element of the initialized particle fitness value set; if the current fitness value is better than the local optimal solution, Step 6: Updating the local optimal solution; if the current fitness value is better than the global optimal solution, updating the global optimal solution; Step 7: Recalculating the velocity and speed of the particles and performing a mutation operation; Step 8: Judging that the number of iterations is less than a preset value, and returning to Step 4; Step 9: Assigning the obtained optimal value to the model for training and learning; Outputting the future energy demand prediction result.

2. The energy demand prediction method based on the ICA algorithm for optimizing the PSO-BP neural network model according to claim 1, wherein The energy impact data includes: total energy, energy usage, climate environment, geographical features, population quantity, economic income, energy pollution degree, and energy price.

3. A method for predicting energy demand based on an ICA algorithm optimized PSO-BP neural network model according to claim 1, characterized in that The preprocessing includes: performing data cleaning on the energy impact data to obtain the first energy impact data; wherein, the data cleaning processes the missing values, outliers, and error values existing in the energy impact data; performing time series processing on the first energy impact data to obtain the second energy impact data; wherein, the time series processing includes: decomposition by date, seasonal adjustment, and trend analysis; the second energy impact data is processed by standardization and normalization to obtain the third energy impact data; performing one-hot encoding on the second energy impact data to obtain the standard energy impact data.

4. A method for predicting energy demand based on an ICA algorithm-optimized PSO-BP neural network model according to claim 1, characterized in that, The energy cost impact correlation matrix is expressed as: ； where EPICM represents the energy cost impact correlation matrix; represents the Kth energy cost impact indicator in the first historical energy impact data; represents the Mth type of energy cost in the first historical energy impact data; represents the Mth type of energy cost in the Nth historical energy impact data.

5. A method for predicting energy demand based on an ICA algorithm-optimized PSO-BP neural network model according to claim 1, characterized in that The calculation formula for the loss value is: ； Among them, N represents the total number of historical energy impact data; M represents the total number of energy categories; represents the predicted cost of the m-th type of energy; represents the actual cost of the m-th type of energy.

6. A method for predicting energy demand based on an ICA algorithm-optimized PSO-BP neural network model according to claim 1, characterized in that The specific implementation of the energy cost model includes: Performing normalization processing on the historical energy cost data for N years to obtain normalized historical energy cost data; Performing clustering on the normalized historical energy cost data to obtain C clustering sets; wherein, C is a natural number greater than 1; Construct C relevant energy data difference matrices for the C clustering sets; among them, the relevant energy data difference matrix is expressed as: ; among them, represents the relevant energy data difference matrix of the i-th clustering set; represents the difference value of the Q-th data in the i-th clustering set with respect to the K-th factor; Based on the clustering set, an energy cost influencing factor transfer matrix is established; wherein, the energy cost influencing factor transfer matrix is expressed as: ; wherein, represents the energy cost influencing factor transfer matrix of the i-th clustering set; represents the transfer probability value between the first energy influencing factor and the K-th energy influencing factor in the i-th clustering set; The constraint conditions of the energy cost influencing factor transfer matrix are as follows: ; Updating the influence factor weight vector according to the energy cost influence factor transfer matrix; wherein, the update formula for the influence factor weight vector is: ； Among them, represents the weight vector of the cost influencing factors of the $i$-th clustering set at time $t$; represents the correction term at time $t$; Establishing an energy cost prediction function; wherein, the formula for the energy cost prediction function is: ； Among them, represents the energy cost prediction function for the i-th clustering set; represents the basic cost of the M-th type of energy; represents the non-linear cost prediction function part; B represents the bias of the energy cost prediction function; Among them, the update formula for the bias is as follows: ; is expressed as a correction parameter; Minimizing the energy cost prediction function; wherein, the minimization process is: ； Among them, T represents the maximum number of iterative training rounds; represents the prediction result of the m-th type of energy in the t-th round by the energy cost prediction function; Performing clustering on the future energy impact data and the clustering sets; obtaining a clustering result; Inputting the future energy impact data into the energy cost prediction function of the corresponding clustering set to obtain the future energy predicted cost.

7. A method for predicting energy demand based on an ICA algorithm-optimized PSO-BP neural network model according to claim 1, characterized in that, The specific process for obtaining the energy development policy features includes: Obtaining relevant energy development policy documents issued in the region; Extracting the energy-related content in the relevant energy development policy documents; Wherein, the energy-related content includes: energy status information, energy planning information, environmental impact information, economic fluctuation information, and energy support information; Adopting a feature engineering method for the energy-related content to obtain the energy development policy feature vector.

8. A method for predicting energy demand based on an ICA algorithm-optimized PSO-BP neural network model according to claim 1, characterized in that The specific process of the energy demand optimization particle swarm algorithm includes: Setting E particle swarms; wherein, E is a natural number greater than 1; Initialize the position set and velocity set of the particles; among them, the position set is expressed as ; which represents the position of the D-th particle in the i-th particle swarm; the velocity set is expressed as ; which represents the velocity of the D-th particle in the i-th particle swarm; Updating the position set and the velocity set of the particles; wherein, the update formula is: ； ； Among them, represents the velocity of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration; represents the velocity of the D-th particle in the i-th particle swarm at the k-th iteration; represents the position of the D-th particle in the i-th particle swarm at the k-th iteration; and represents the learning factor as a non-negative constant; represents the local optimal value of the D-th particle in the i-th particle swarm; represents the global optimal value of the D-th particle; represents at the k-th iteration random number; represents the inertia weight coefficient; represents the position of the D-th particle in the i-th particle swarm after the (k + 1)-th iteration; Wherein, the calculation formula for the inertia weight coefficient is: ； Among them, t represents the number of iterations; represents the maximum value of the inertia weight coefficient; represents the minimum value of the inertia weight coefficient; represents the maximum number of iterations; d represents the initial inertia weight after the initial search.

Images