A calculation method for energy demand forecasting based on substitution model optimization

Abstract

A method for calculating energy demand forecasting based on alternative model optimization belongs to the technical field of energy demand forecasting. In order to solve the impact of complex nonlinear relationships on energy demand forecasting, the present invention includes constructing an initial model for energy demand forecasting based on temperature, industrial output value, and business activity index; constructing a nonlinear gradient decomposition intermediate function based on nested transformation and nonlinear mapping operations; generating intermediate variables using weighted integral transformation; then performing a second transition operation based on the intermediate variables to construct a composite adjustment factor; constructing an energy demand forecast alternative model; and using the prediction results to evaluate the model. The strategy of dynamically selecting alternative models proposed in the present invention can flexibly switch between alternative models and initial models according to the needs of actual problems. When the objective function is complex or the accuracy is insufficient, the alternative model can improve the data dimension by mining internal management and provide more new information that does not require additional acquisition. The present invention improves computing efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy demand forecasting, and in particular relates to an energy demand forecasting calculation method based on substitution model optimization. Background Art

[0002] A stable supply and rational allocation of energy are crucial to the normal operation and development of cities. Accurately predicting hourly energy demand has become a critical issue that needs to be addressed in energy management. Accurate hourly energy demand forecasts help energy suppliers plan production and distribution in advance, avoid energy waste or shortages, ensure the stability of urban energy supply, reduce energy costs, and mitigate economic losses caused by insufficient or excessive energy supply. Furthermore, from an environmental perspective, reasonable energy forecasts can help optimize the energy structure, promote sustainable energy utilization, reduce carbon emissions, and achieve green development goals. Therefore, achieving high-precision hourly energy demand forecasts is of great significance in many economic, social, and environmental contexts, and its urgency is self-evident.

[0003] Machine learning models have become the mainstream approach in energy demand forecasting. Many machine learning-based prediction models rely on large amounts of data for training. When sufficient data is available, these models leverage complex algorithms and powerful computing power to uncover underlying patterns and regularities within the data, thereby providing relatively accurate forecasts. However, these approaches have significant limitations. When data volumes are insufficient, model performance deteriorates significantly, and prediction accuracy plummets, failing to meet practical needs. Furthermore, in some real-world scenarios, obtaining large amounts of high-quality data is challenging. High data collection costs, privacy concerns, and the difficulty of acquiring data all hinder the application of these machine learning models that rely on large amounts of data.

[0004] In contrast, traditional energy demand forecasting methods have been widely questioned in the era of big data. Although traditional methods have faced considerable scrutiny in the data-intensive sector, they still demonstrate certain advantages when dealing with smaller amounts of data. Traditional methods are typically based on simple statistical analysis or empirical models, do not require extensive data support, and have relatively simple calculations. This allows them to quickly generate forecasts and achieve high efficiency when processing small amounts of data. However, traditional methods have relatively little room for improvement. When faced with complex energy demand scenarios, especially those with complex nonlinear relationships, they struggle to accurately capture the inherent connections between various factors, making it difficult to improve forecast accuracy. Summary of the Invention

[0005] The problem to be solved by the present invention is to overcome the limitations of data volume and data dimension, while effectively dealing with the impact of complex nonlinear relationships on energy demand forecasting, and proposing an energy demand forecasting calculation method based on substitution model optimization.

[0006] To achieve the above object, the present invention is implemented through the following technical solutions:

[0007] A method for calculating energy demand forecasting based on substitution model optimization includes the following steps:

[0008] S1. Collect temperature data, industrial output data, business activity index data, and corresponding energy demand data at each time point to construct a historical data set;

[0009] S2. Construct an initial energy demand forecast model based on temperature, industrial output value, and business activity index to conduct energy demand forecasting;

[0010] S3. For the initial energy demand forecast model obtained in step S2, construct a nonlinear gradient decomposition intermediate function based on nested transformation and nonlinear mapping operation;

[0011] S4. The nonlinear gradient decomposition intermediate function obtained in step S3 is decomposed by a weighted integral transform to generate an intermediate variable; a second transition operation is then performed based on the intermediate variable to construct a composite adjustment factor;

[0012] S5. Based on the composite adjustment factor constructed in step S4, the initial energy demand forecast model obtained in step S2 is adjusted to obtain an alternative energy demand forecast model for energy demand forecasting;

[0013] S6. Use the prediction results obtained in steps S2 and S5 to perform model evaluation.

[0014] Furthermore, the industrial output value data in step S1 is the monthly output value data of industrial enterprises in the region; the business activity index data is calculated by counting the flow of people in shopping malls and office buildings and combining it with the business turnover data.

[0015] Furthermore, the energy demand forecasting initial model f(NENGY i ) is:

[0016] f(NENGY i )=αWEND i +βGONGY i +δSHANGY i +ε

[0017] Among them, WEND i is the temperature data at the i-th time point, GONGY is the industrial output value data at the i-th time point, SHANGY iis the business activity index data at the i-th time point, α, β, and δ are the fitting coefficients of the temperature data at the i-th time point, the industrial output value data at the i-th time point, and the business activity index data at the i-th time point, respectively, and ε is the error term;

[0018] The fitting coefficients α, β, and δ are calculated based on the historical data set, and then the predicted value at each i time point is output as D i , calculate the difference between the predicted value and the true value and record it as e i , the expression is:

[0019] e i =NENGY i -(αWEND i +βGONGY i +δSHANGY i ).

[0020] Furthermore, in step S3, a nonlinear gradient decomposition intermediate function H is constructed based on nested transformation and nonlinear mapping operation. i The expression is:

[0021]

[0022] dl i =αWEND i +βGONGY i +δSHANGY i -θ ijp

[0023] Among them, p represents the dimension when the sigmoid function is performed, q represents the total number of dimensions when the sigmoid function is performed; j represents the number parameter when the sigmoid function is performed, m represents the total number of number parameters when the sigmoid function is performed; w ijp is the weight coefficient determined by considering the function operation dimension and quantity parameters at each i-th time point according to the historical fluctuation law; sgn is the sign function; dl i is the agent intermediate process function; θ ijp is the threshold value set at each i-th time point based on the historical fluctuation law and considering the function operation dimension and quantity parameters; σ(·) is the sigmoid function, which is used to introduce nonlinearity; b ijp is the bias term; r ijp It is an exponential parameter used to adjust the weight ratio of each part, which is optimized by cross-validation method; r is the dimension when performing the hyperbolic tangent function operation, s is the number parameter when performing the hyperbolic tangent function operation; u ir is the coefficient associated with the latent characteristics of the data; is a scaling factor determined based on the variance and range of the data; tanh(·) is the hyperbolic tangent function.

[0024] Furthermore, the specific implementation method of step S4 includes the following steps:

[0025] S4.1. Use weighted integral transformation to generate intermediate variable M for the nonlinear gradient decomposition intermediate function obtained in step S3 i , the expression is:

[0026]

[0027] Among them, η ik It is a weight coefficient that depends on the spatiotemporal characteristics of the data points and is used to emphasize the importance of data in different local areas; sinc() is the Sinker function; It is a parameter related to the frequency and phase of data fluctuation; dh is the integral variable in the integral operation. dh is the integration of variable h, which is a dummy variable whose value ranges from 0 to H. i ;

[0028] S4.2. Perform a second transition operation based on the intermediate variables to construct the composite adjustment factor N i , the expression is:

[0029]

[0030] Among them, softplus is a nonlinear function, ζ i ,ρ i are the scaling factor and bias term respectively.

[0031] Furthermore, the specific implementation method of step S5 includes the following steps:

[0032] S5.1. Adjust the fitting coefficient of the initial energy demand forecast model based on the composite adjustment factor constructed in step S4. The expression is:

[0033]

[0034]

[0035] in, are the derivatives of the fitting coefficients α, β, and δ, respectively. Is the gradient operator, which means finding the gradient of the variable;

[0036] S5.2. The calculated With the original WEND i ,GONGY i ,SHANGY iComposed of new data combination (WEND i ,GONGY i ,SHANGY i , ), and then construct the energy demand forecast alternative model f′(NENGY i ), the expression is:

[0037]

[0038] Among them, α′, β′, δ′ are the fitting coefficients of the temperature data at the i-th time point, the industrial output value data at the i-th time point, and the business activity index data at the i-th time point after the update. γ, ι, ψ are The corresponding fitting coefficient, ε′ is the error term of the alternative model for energy demand forecasting;

[0039] S5.3. For each NENGY in the historical dataset i The energy demand forecasting alternative model outputs a forecast value, denoted as XD i , the difference between the predicted value and the true value is xe i , the expression is:

[0040]

[0041] Furthermore, the specific implementation method of step S6 includes the following steps:

[0042] S6.1. For the i-th time point, calculate the error η between the prediction values ​​of the alternative energy demand forecast model and the initial energy demand forecast model. i , the expression is:

[0043]

[0044] Among them, L 2 is the norm;

[0045] S6.2. Set the comparison threshold, and then calculate the η i Perform analysis:

[0046] When η i If the value is less than the comparison threshold, it is determined that the performance of the energy demand forecast alternative model and the initial energy demand forecast model meets the requirements. In this case, the linear effect in the initial energy demand forecast model is used to explain the influence coefficient of each variable on energy demand. α, β, and δ are directly used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted value of the initial energy demand forecast model.

[0047] When η i is greater than the comparison threshold and |(xei -e i ) / e i |≥0.001, the linear effect in the energy demand forecast alternative model is used to explain the impact coefficient of each variable on energy demand, and the fitting coefficients α′, β′, and δ′ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted values ​​of the energy demand forecast alternative model;

[0048] When η i is greater than the comparison threshold and |(xe i -e i ) / e i |<0.001, it is judged that the linear impact in the initial energy demand forecast model is used to explain the impact coefficient of each variable on energy demand, and the fitting coefficients α, β, and δ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted values ​​of the initial energy demand forecast model.

[0049] Furthermore, in step S6, the comparison threshold is set to 0.1.

[0050] Beneficial effects of the present invention:

[0051] The energy demand forecasting calculation method based on substitution model optimization described in the present invention integrates multiple technical advantages, aims to take into account the forecasting needs under different data volumes, deeply explore the potential information in the data dimensions, and fully consider the complex nonlinear relationship between energy demand and various influencing factors, thereby providing a more accurate and efficient solution for energy demand forecasting.

[0052] The energy demand forecasting calculation method based on substitution model optimization described in the present invention deeply explores the complex relationship between the three variables of temperature, industrial output value and business activity index by constructing intermediate functions and applying weighted integral transformation, so as to improve the accuracy and reliability of the forecast and provide a more effective solution for energy demand forecasting.

[0053] The energy demand forecasting calculation method based on surrogate model optimization described in this paper significantly improves optimization computational efficiency. By implementing a derivative-based optimization method (gradient descent) as a replacement during the optimization process, it maintains high efficiency, especially when data is insufficient. This method offers invaluable computational advantages, particularly in real-time systems or high-frequency iterative optimization problems.

[0054] The energy demand forecasting and calculation method based on surrogate model optimization described in this invention reduces computational costs while ensuring high accuracy. By rationally selecting surrogate models, particularly utilizing smooth and differentiable proxy models, the number of direct objective function calculations can be reduced compared to machine learning models, while still ensuring optimization accuracy. This is particularly important for computationally expensive objective functions, as it avoids frequent calculations of the actual objective function, thereby saving significant computing resources. Through error control and model selection strategies, the optimization results can meet accuracy requirements while reducing resource consumption.

[0055] The energy demand forecasting calculation method based on surrogate model optimization described in this invention improves the stability and convergence speed of optimization results. Traditional derivative-free optimization methods may suffer from slow convergence or unstable results due to the lack of derivative information. By introducing a surrogate model, the optimization process can more effectively utilize gradient information, significantly improving the convergence speed. In a large number of experiments and applications, the optimization process has demonstrated more stable convergence and higher final accuracy than derivative-free methods, especially in complex high-dimensional optimization problems.

[0056] The energy demand forecasting and calculation method based on surrogate model optimization described in the present invention enhances the adaptability and flexibility of the optimization method. The strategy for dynamically selecting surrogate models proposed in the present invention allows for flexible switching between surrogate models and the initial model based on the needs of the actual problem. When the objective function is complex or lacks precision, the surrogate model can enhance data dimensionality by mining internal management, providing more new information that does not require additional acquisition, thereby achieving optimal computational efficiency and optimization effects in different scenarios. This flexibility makes the present invention applicable to a wider range of optimization applications.

[0057] The energy demand forecasting and calculation method based on surrogate model optimization described in the present invention has broad application prospects. The present invention is not limited to a single field; its method can be widely applied to multiple industries and fields, including but not limited to machine learning model training, financial optimization, intelligent manufacturing, and the like. In particular, in real-time optimization and complex multi-objective optimization problems, the technical solution of the present invention can significantly improve the optimization effect and shorten the optimization cycle, and has broad application prospects. While ensuring optimization accuracy, it significantly improves computational efficiency, has significant economic and technical advantages, and can provide a practical solution for practical engineering and industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Picture 1 This is a flow chart of an energy demand forecasting calculation method based on substitution model optimization described in the present invention. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present invention and are not intended to limit the present invention. That is, the specific embodiments described herein are only some embodiments of the present invention, not all embodiments. Generally, the components of the specific embodiments of the present invention described and illustrated in the drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0060] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected specific embodiments of the present invention. All other specific embodiments obtained by those skilled in the art based on the specific embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0061] In order to further understand the content, features and effects of the present invention, the following specific embodiments are given as examples, and the attached Picture 1 The detailed instructions are as follows:

[0062] Example 1:

[0063] A method for calculating energy demand forecasting based on substitution model optimization includes the following steps:

[0064] S1. Collect temperature data, industrial output data, business activity index data, and corresponding energy demand data at each time point to construct a historical data set;

[0065] Furthermore, the industrial output value data in step S1 is the monthly output value data of industrial enterprises in the region; the commercial activity index data is calculated by counting the flow of people in shopping malls and office buildings and combining it with the business turnover data;

[0066] Furthermore, daily and hourly temperature data for at least one year are collected, covering temperature changes in different seasons and weather conditions. The data source can be public data from local weather stations or professional meteorological data platforms.

[0067] Furthermore, industrial output data can be obtained from local statistical bureau websites, industrial enterprise report summaries, etc. For small enterprises that cannot directly obtain output data, approximate estimates can be made based on their electricity consumption, production scale, etc.

[0068] Furthermore, the above data is organized into a table format, where each row represents a record at a time point, and each column represents temperature, industrial output value, business activity index, time, and corresponding energy demand. Some of the collected data are as follows:

[0069] No WEND GONGY SHANGY NENGY No WEND GONGY SHANGY NENGY 1 -12.8437 114.4739 107.4624 868.462 21 -3.82246 105.9528 98.41308 823.7406 2 -10.2043 112.6742 106.7384 899.3065 22 -3.43742 120.6546 83.25045 802.7143 3 -7.05928 149.2999 101.1501 860.3098 23 -3.06888 130.3089 91.99783 846.6879 4 -6.99385 147.8133 88.72375 877.9371 24 -2.72641 152.8263 97.2172 835.4771 5 -6.22295 144.6273 86.14398 890.3436 25 -2.52869 120.6813 95.18869 796.7444 6 -5.79028 126.7239 78.0911 841.6784 26 -2.48793 139.0303 62.57173 777.106 7 -5.71612 134.8482 72.26702 809.5493 27 -2.36433 101.9337 83.91096 806.4836 8 -5.68686 124.5118 86.4957 857.7312 28 -2.32984 116.4532 92.04511 783.5975 9 -5.67252 113.3909 99.21264 869.4055 29 -2.32223 133.608 94.5222 814.2649 10 -5.19819 150.9717 69.75644 840.6552 30 -2.29271 109.3687 71.22736 789.8584 11 -5.05151 153.1235 90.28054 865.3869 31 -2.19028 122.3096 84.80708 752.5231 12 -4.83333 127.1611 83.69705 822.6742 32 -1.90622 132.9546 76.5097 817.7687 13 -4.82097 146.2367 83.44231 828.1119 33 -1.76236 134.27 78.38026 810.832 14 -4.66352 118.3629 102.0156 808.9585 34 -1.35128 135.073 71.49636 767.8903 15 -4.51799 114.3892 97.72461 839.5121 35 -0.5322 148.5888 92.49482 800.4536 16 -4.27222 115.8124 122.7773 855.3487 36 -0.49306 134.3463 93.49233 816.9448 17 -4.25874 149.6863 76.88267 824.7523 37 -0.4818 152.5715 105.9624 812.1229 18 -4.1693 154.6424 92.54768 860.3141 38 -0.34231 142.1526 86.33775 804.0109 19 -4.02191 122.5094 95.62311 801.3009 39 -0.3107 141.8373 85.64953 799.6461 20 -3.9261 149.6797 83.7328 857.5021 40 … … … …

[0070] Furthermore, the above data are organized into a table format, where each row represents a record at a time point, and each column is temperature, industrial output value, business activity index, time, and corresponding energy demand.

[0071] S2. Construct an initial energy demand forecast model based on temperature, industrial output value, and business activity index to conduct energy demand forecasting;

[0072] Furthermore, the energy demand forecasting initial model f(NENGY i ) is:

[0073] f(NENGY i )=αWEND i +βGONGY i +δSHANGY i +ε

[0074] Among them, WEND i is the temperature data at the i-th time point, GONGY is the industrial output value data at the i-th time point, SHANGY i is the business activity index data at the i-th time point, α, β, and δ are the fitting coefficients of the temperature data at the i-th time point, the industrial output value data at the i-th time point, and the business activity index data at the i-th time point, respectively, and ε is the error term;

[0075] Furthermore, the fitting coefficients α, β, and δ calculated based on the historical data set are -2.403709069, 0.923278621, and 0.717843077, respectively.

[0076] The fitting coefficients α, β, and δ are calculated based on the historical data set, and then the predicted value at each i time point is output as D i , calculate the difference between the predicted value and the true value and record it as e i , the expression is:

[0077] e i =NENGY i -(αWEND i +βGONGY i +δSHANGY i ).

[0078] S3. For the initial energy demand forecast model obtained in step S2, construct a nonlinear gradient decomposition intermediate function based on nested transformation and nonlinear mapping operation;

[0079] Furthermore, in step S3, a nonlinear gradient decomposition intermediate function H is constructed based on nested transformation and nonlinear mapping operation. i The expression is:

[0080]

[0081] dl i =αWEND i +βGONGY i +δSHANGY i -θ ijp

[0082] Among them, p represents the dimension when the sigmoid function is performed, q represents the total number of dimensions when the sigmoid function is performed; j represents the number parameter when the sigmoid function is performed, m represents the total number of number parameters when the sigmoid function is performed; w ijp is the weight coefficient determined by considering the function operation dimension and quantity parameters at each i-th time point according to the historical fluctuation law; sgn is the sign function; dl i is the agent intermediate process function; θ ijp is the threshold value set at each i-th time point based on the historical fluctuation law and considering the function operation dimension and quantity parameters; σ(·) is the sigmoid function, which is used to introduce nonlinearity; b ijp is the bias term; r ijp It is an exponential parameter used to adjust the weight ratio of each part, which is optimized by cross-validation method; r is the dimension when performing the hyperbolic tangent function operation, s is the number parameter when performing the hyperbolic tangent function operation; u ir is the coefficient associated with the latent characteristics of the data; is a scaling factor determined based on the variance and range of the data; tanh(·) is the hyperbolic tangent function;

[0083] Gradient calculation is crucial when optimizing a surrogate model. It accurately reveals the direction and rate of change of the objective function at the current point, which is crucial for guiding the algorithm iteratively toward the optimal solution. Obtaining partial derivatives with respect to the model coefficients requires constructing a complex and effective intermediate function. To ensure that this intermediate function fully captures the complex relationships in the data, we employ a series of carefully designed computational steps.

[0084] First, weighted operations are performed. Different data features have varying degrees of influence on energy demand. Temperature data has varying impacts on energy demand in different seasons, and industrial output data can also fluctuate due to factors such as industrial restructuring and policy changes. Therefore, weight coefficients are introduced. These coefficients are not fixed but take into account multiple factors, including data characteristics, historical fluctuations, and the impact of different time periods. By weighting different features, the importance of key factors in the calculation is highlighted, making the intermediate function more accurate to actual energy demand fluctuations. Nested transformations further deepen the exploration of data relationships. In real-world energy demand scenarios, various factors are not simply linearly correlated but rather intertwined and nested. The impact of temperature on energy demand is modulated by industrial production activities, which in turn influence industrial energy consumption patterns to a certain extent. Second, to model this complex nested relationship, a multi-layered operation structure is employed in the intermediate function. For example, the sigmoid function maps input data to the range 0 to 1, introducing nonlinearity while also filtering and adjusting combinations of different features. Each nested layer in the function acts like a filter, gradually extracting the most representative information from the data, thereby more accurately reflecting the combined impact of various factors on energy demand. Nonlinear mapping is also an important method for constructing intermediate functions. The relationship between energy demand and various influencing factors is often nonlinear, making it difficult for simple linear models to accurately depict it.

[0085] The use of the hyperbolic tangent and sigmoid functions can transform linear input data nonlinearly, expanding the expressive power of the functions. In extreme high or low temperatures, the impact of temperature on energy demand does not increase linearly. By mapping these nonlinear functions, we can more realistically simulate this complex trend. Furthermore, squaring the data and dividing it by a scaling factor can further adjust the data's distribution and magnitude, allowing the intermediate functions to better adapt to the diversity and complexity of the data.

[0086] S4. The nonlinear gradient decomposition intermediate function obtained in step S3 is decomposed by a weighted integral transform to generate an intermediate variable; a second transition operation is then performed based on the intermediate variable to construct a composite adjustment factor;

[0087] Furthermore, the specific implementation method of step S4 includes the following steps:

[0088] S4.1. Use weighted integral transformation to generate intermediate variable M for the nonlinear gradient decomposition intermediate function obtained in step S3 i , the expression is:

[0089]

[0090] Among them, η ik It is a weight coefficient that depends on the spatiotemporal characteristics of the data points and is used to emphasize the importance of data in different local areas; sinc() is the Sinker function; It is a parameter related to the frequency and phase of data fluctuation; dh is the integral variable in the integral operation. In the example, the variable h is integrated. h is a dummy variable whose value ranges from 0 to H. i ;

[0091] Furthermore, the above operation can convert H i The information contained in the data is redistributed and integrated at different frequencies and local areas to provide more representative data features for subsequent calculations;

[0092] S4.2. Perform a second transition operation based on the intermediate variables to construct the composite adjustment factor N i , the expression is:

[0093]

[0094] Among them, softplus is a nonlinear function, ζ i ,ρ i are the scaling factor and bias term respectively.

[0095] softplus(x)=ln(1+e x ) is a smooth nonlinear function similar to the ReLU function, which can map the input data to the positive domain while maintaining certain nonlinear characteristics; i ,ρ i Through comparative analysis of different data subsets and sensitivity testing of the model, the optimization is aimed at i The value of dynamically adjusts its impact on the final calculation, enhancing the model's adaptability to complex data patterns.

[0096] Furthermore, to accurately calculate the partial derivatives of the fitting coefficients, it is necessary to further explore the potential information correlations within this intermediate function. This transition operation uses a weighted integral transform to generate new intermediate variables. The weighted integral transform is a powerful mathematical technique that can comprehensively analyze the intermediate function from multiple dimensions and scales. This operation also utilizes a special function that cleverly processes the intermediate function in the frequency domain. It selects information components at different frequencies based on parameters related to the frequency of data fluctuations.

[0097] S5. Based on the composite adjustment factor constructed in step S4, the initial energy demand forecast model obtained in step S2 is adjusted to obtain an alternative energy demand forecast model for energy demand forecasting;

[0098] Furthermore, the specific implementation method of step S5 includes the following steps:

[0099] S5.1. Adjust the fitting coefficient of the initial energy demand forecast model based on the composite adjustment factor constructed in step S4. The expression is:

[0100]

[0101]

[0102] in, are the derivatives of the fitting coefficients α, β, and δ, respectively. Is the gradient operator, which means finding the gradient of the variable;

[0103] tanh(|e i The function of |) is to dynamically adjust the weight of the data point in the gradient calculation according to the deviation between the actual value and the predicted value;

[0104] S5.2. The calculated With the original WEND i ,GONGY i ,SHANGY i Composed of new data combination (WEND i ,GONGY i ,SHANGY i , ), and then construct the energy demand forecast alternative model f′(NENGY i ), the expression is:

[0105]

[0106] Among them, α′, β′, δ′ are the fitting coefficients of the temperature data at the i-th time point, the industrial output value data at the i-th time point, and the business activity index data at the i-th time point after the update. γ, ι, ψ are The corresponding fitting coefficient, ε′ is the error term of the alternative model for energy demand forecasting;

[0107] Furthermore, compared with the initial model fitting coefficients α, β, δ (-2.403709069, 0.923278621 and 0.717843077), α′, β′, δ′ are updated to -2.6635458522, 0.963254552 and 0.74412658.

[0108] S5.3. For each NENGY in the historical dataset iThe energy demand forecasting alternative model outputs a forecast value, denoted as XD i , the difference between the predicted value and the true value is xe i , the expression is:

[0109]

[0110] S6. Using the prediction results obtained in steps S2 and S5 to perform model evaluation;

[0111] Furthermore, the specific implementation method of step S6 includes the following steps:

[0112] S6.1. For the i-th time point, calculate the error η between the prediction values ​​of the alternative energy demand forecast model and the initial energy demand forecast model. i , the expression is:

[0113]

[0114] Among them, L 2 is the norm;

[0115] S6.2. Set the comparison threshold, and then calculate the η i Perform analysis:

[0116] When η i If the value is less than the comparison threshold, it is determined that the performance of the energy demand forecast alternative model and the initial energy demand forecast model meets the requirements. In this case, the linear effect in the initial energy demand forecast model is used to explain the influence coefficient of each variable on energy demand. α, β, and δ are directly used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted value of the initial energy demand forecast model.

[0117] When η i is greater than the comparison threshold and |(xe i -e i ) / e i |≥0.001, the linear effect in the energy demand forecast alternative model is used to explain the impact coefficient of each variable on energy demand, and the fitting coefficients α′, β′, and δ′ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted values ​​of the energy demand forecast alternative model;

[0118] When η i is greater than the comparison threshold and |(xe i -e i ) / e i|<0.001, it is judged that the linear impact in the initial energy demand forecast model is used to explain the impact coefficient of each variable on energy demand, and the fitting coefficients α, β, and δ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted values ​​of the initial energy demand forecast model.

[0119] Furthermore, in step S6, the comparison threshold is set to 0.1.

[0120] Take 13 sets of data as an example, among these 13 sets of data, η i They are 0.211, 0.255, 0.032, 0.445, 0.365, 0.326, 0.125, 0.123, 0.022, 0.556, 0.114, 0.556, and 0.336 respectively. i -e i ) / e i The value ratio is 0.0022, 0.0063, 0.0045, 0.0055, 0.0033, 0.0007, 0.0067, 0.0021, 0.0021, 0.0055, 0.0016, 0.0004, and 0.0096.

[0121] The η values ​​of the data of group 3 (0.032) and group 9 (0.022) were i If the value is less than the comparison threshold of 0.1 (marked with a gray background in the system), the linear effect in the initial energy demand forecast model is used to explain the impact coefficient of each variable on energy demand, and α, β, and δ are directly used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand.

[0122] In the data of group 6 (0.0007) and group 12 (0.0004), |(xe i -e i ) / e i |<0.001 (marked with blue background in the system), in this case, α, β, and δ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand; in other cases, the fitting coefficients α′, β′, and δ′ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand.

[0123] The key technical points and points to be protected of the present invention are: a complete model framework and the mathematical expression of the core algorithm.

[0124] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0125] Although the present application has been described above with reference to specific embodiments, various modifications may be made thereto and components may be substituted with equivalents without departing from the scope of the present application. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of these combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions within the scope of the claims.

Claims

1. A method for energy demand forecasting based on substitution model optimization, characterized in that: The steps include: S1. Collect temperature data, industrial output data, business activity index data, and corresponding energy demand data at each time point to construct a historical data set; S2. Construct an initial energy demand forecast model based on temperature, industrial output value, and business activity index to conduct energy demand forecasting; S3. For the initial energy demand forecast model obtained in step S2, construct a nonlinear gradient decomposition intermediate function based on nested transformation and nonlinear mapping operation; In step S3, a nonlinear gradient decomposition intermediate function H is constructed based on nested transformation and nonlinear mapping operation. i The expression is: d.l. i =αWEND i +βGONGY i +δSHANGY i -θ ijp Among them, p represents the dimension when the sigmoid function is performed, q represents the total number of dimensions when the sigmoid function is performed; j represents the number parameter when the sigmoid function is performed, m represents the total number of number parameters when the sigmoid function is performed; w ijp is the weight coefficient determined by considering the function operation dimension and quantity parameters at each i-th time point according to the historical fluctuation law; sgn is the sign function; dl i is the agent intermediate process function; θ ijp is the threshold value set at each i-th time point based on the historical fluctuation law and considering the function operation dimension and quantity parameters; σ(·) is the sigmoid function, which is used to introduce nonlinearity; b ijp is the bias term; r ijp It is an exponential parameter used to adjust the weight ratio of each part, which is optimized by cross-validation method; r is the dimension when performing the hyperbolic tangent function operation, s is the number parameter when performing the hyperbolic tangent function operation; u ir is the coefficient associated with the latent characteristics of the data; is a scaling factor determined based on the variance and range of the data; tanh(·) is the hyperbolic tangent function; WEND i is the temperature data at the i-th time point, GONGY is the industrial output value data at the i-th time point, SHANGY i is the business activity index data at the i-th time point, α, β, and δ are the fitting coefficients of the temperature data at the i-th time point, the industrial output value data at the i-th time point, and the business activity index data at the i-th time point, respectively; S4. The nonlinear gradient decomposition intermediate function obtained in step S3 is decomposed by a weighted integral transform to generate an intermediate variable; a second transition operation is then performed based on the intermediate variable to construct a composite adjustment factor; S5. Based on the composite adjustment factor constructed in step S4, the initial energy demand forecast model obtained in step S2 is adjusted to obtain an alternative energy demand forecast model for energy demand forecasting; S6. Use the prediction results obtained in steps S2 and S5 to perform model evaluation.

2. The energy demand forecasting calculation method based on substitution model optimization according to claim 1 is characterized in that: The industrial output value data in step S1 is the monthly output value data of industrial enterprises in the region; the business activity index data is calculated by counting the flow of people in shopping malls and office buildings and combining it with the business turnover data.

3. The energy demand forecasting calculation method based on substitution model optimization according to claim 2 is characterized in that: The energy demand forecasting initial model f(NENGY i ) is: f(NENGY i )=aWEND i +βGONGY i +δSHANGY i +e Among them, WEND i is the temperature data at the i-th time point, GONGY is the industrial output value data at the i-th time point, SHANGY i is the business activity index data at the i-th time point, α, β, and δ are the fitting coefficients of the temperature data at the i-th time point, the industrial output value data at the i-th time point, and the business activity index data at the i-th time point, respectively, and ε is the error term; The fitting coefficients α, β, and δ are calculated based on the historical data set, and then the predicted value at each i time point is output as D i , calculate the difference between the predicted value and the true value and record it as e i , the expression is: e i =NENGY i -(αWEND i +βGONGY i +δSHANGY i )。 4. The energy demand forecasting calculation method based on substitution model optimization according to claim 3 is characterized in that: The specific implementation method of step S4 includes the following steps: S4.

1. Use weighted integral transformation to generate intermediate variable M for the nonlinear gradient decomposition intermediate function obtained in step S3 i , the expression is: Among them, η ik It is a weight coefficient that depends on the spatiotemporal characteristics of the data points and is used to emphasize the importance of data in different local areas; sinc() is the Sinker function; It is a parameter related to the frequency and phase of data fluctuation; dh is the integral variable in the integral operation. In the example, the variable h is integrated. h is a dummy variable whose value ranges from 0 to H. i ; S4.

2. Perform a second transition operation based on the intermediate variables to construct the composite adjustment factor N i , the expression is: Among them, softplus is a nonlinear function, ζ i ,ρ i are the scaling factor and bias term respectively.

5. The energy demand forecasting calculation method based on substitution model optimization according to claim 4 is characterized in that: The specific implementation method of step S5 includes the following steps: S5.

1. Adjust the fitting coefficient of the initial energy demand forecast model based on the composite adjustment factor constructed in step S4. The expression is: in, are the derivatives of the fitting coefficients α, β, and δ, respectively. Is the gradient operator, which means finding the gradient of the variable; S5.

2. The calculated With the original WEND i ,GONGY i ,SHANGY i Composed of new data combination (WEND i ,GONGY i ,SHANGY i , ), and then construct the energy demand forecast alternative model f′(NENGY i ), the expression is: Among them, α′, β′, δ′ are the fitting coefficients of the temperature data at the i-th time point, the industrial output value data at the i-th time point, and the business activity index data at the i-th time point after the update. γ, ι, ψ are The corresponding fitting coefficient, ε′ is the error term of the alternative model for energy demand forecasting; S5.

3. For each NENGY in the historical dataset i The energy demand forecasting alternative model outputs a forecast value, denoted as XD i , the difference between the predicted value and the true value is xe i , the expression is:

6. The energy demand forecasting calculation method based on substitution model optimization according to claim 5 is characterized in that: The specific implementation method of step S6 includes the following steps: S6.

1. For the i-th time point, calculate the error η between the prediction values ​​of the alternative energy demand forecast model and the initial energy demand forecast model. i , the expression is: Among them, L 2 is the norm; S6.

2. Set the comparison threshold, and then calculate the η i Perform analysis: When η i If the value is less than the comparison threshold, it is determined that the performance of the energy demand forecast alternative model and the initial energy demand forecast model meets the requirements. In this case, the linear effect in the initial energy demand forecast model is used to explain the influence coefficient of each variable on energy demand. α, β, and δ are directly used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted value of the initial energy demand forecast model. When η i is greater than the comparison threshold and |(xe i -e i ) / e i |≥0.001, the linear effect in the energy demand forecast alternative model is used to explain the impact coefficient of each variable on energy demand, and the fitting coefficients α′, β′, and δ′ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted values ​​of the energy demand forecast alternative model; When η i is greater than the comparison threshold and |(xe i -e i ) / e i |<0.001, it is judged that the linear impact in the initial energy demand forecast model is used to explain the impact coefficient of each variable on energy demand, and the fitting coefficients α, β, and δ are used to explain the contribution of each unit increase in temperature, industrial and commercial output value to energy demand. Subsequent forecasts use the predicted values ​​of the initial energy demand forecast model.

7. The energy demand forecasting calculation method based on substitution model optimization according to claim 6 is characterized in that: In step S6, the comparison threshold is set to 0.1.

Images