Thermal load prediction method and system based on multi-stage fusion optimization

Abstract

The invention discloses a thermal load prediction method and system based on multi-stage fusion optimization, and the method employs a gray wolf optimization algorithm to replace a conventional empirical method to optimize key parameters of variational mode decomposition, and automatically determines an optimal mode number and a penalty factor through multi-objective optimization. The obtained parameter combination is introduced into a variational mode decomposition algorithm, and then the variational mode decomposition algorithm is utilized to decompose thermal load historical data into a plurality of stable mode components, so that the complexity and noise interference of the data are reduced, and the decomposition quality of variational mode decomposition is remarkably improved; then, the hyper-parameters of the long-short-term memory network are optimized by combining a sparrow search algorithm, so that an optimal hyper-parameter combination is found, and finally, a thermal load data set is predicted through a long-short-term memory model under the obtained optimal hyper-parameter combination.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent prediction of energy systems, and particularly relates to a heat load prediction method and system based on multi-stage fusion optimization. Background Art

[0002] With the growth of global energy demand and the prominence of environmental problems, improving energy utilization efficiency and optimizing the energy structure have become common global challenges. The urban heating system urgently needs to transform towards low carbon, implement a heating strategy with green and low-carbon energy as the main body, and gradually replace the traditional coal-fired heating method. Through the concept of intelligent heating, artificial intelligence and optimization algorithms are used to improve the operation efficiency of heat exchange stations, avoid overheating, reduce energy waste, and promote the green and low-carbon development of the heating system. As a key link in energy management and dispatching, the accuracy of heat load prediction directly affects the energy supply-demand balance, system operation efficiency and economic benefits, and is of great significance to the stability, economy and environmental protection of the heating system.

[0003] Traditional heat load prediction methods mainly rely on statistical models, machine learning algorithms and deep learning techniques, etc. Statistical models conduct predictions by analyzing historical data and establishing mathematical models. For example, some classic statistical models and their variants such as Multiple Linear Regression (MLR), Autoregression Moving Average (ARMA), and Seasonal Autoregressive Integrated Moving Average (SARIMA) have been applied to the field of heat load prediction. Machine learning algorithms are based on a large number of training samples and use various algorithms to establish models, such as Support Vector Machine (SVM), Random Forest (RF), XGBoost, and Artificial Neural Network (ANN).

[0004] In addition, with the rapid development of artificial intelligence technology, especially deep learning technology, which has made breakthrough progress in the field of artificial intelligence in recent years. It simulates the working mode of the human brain by constructing a multi-layer neural network and improves the accuracy and generalization ability of the model through training with a large-scale dataset. In the energy field, deep learning has been widely applied to solve various problems, including energy load forecasting, energy consumption optimization, and energy equipment fault detection. Deep learning is effective in solving these problems and has become the mainstream model for short-term heat load forecasting in recent years. In existing models based on a single optimization algorithm (such as SSA-LSTM) or a dual algorithm (such as VMD-SSA-LSTM), it is easy to fall into local optimum and the model stability is poor. To solve this problem, a heat load forecasting method and system based on multi-stage fusion optimization are proposed to achieve multi-level and full-process optimization of load data and deep modeling. Summary of the Invention

[0005] The purpose of the present invention is to provide a heat load forecasting method and system based on multi-stage fusion optimization.

[0006] In the first aspect, the present invention provides a heat load forecasting method based on multi-stage fusion optimization, which includes the following steps: Obtain the historical heat load data and meteorological data of the measured integrated energy system; perform parameter optimization on the penalty factor and the number of mode decompositions of variational mode decomposition, and according to the optimal penalty factor and the number of mode decompositions, use variational mode decomposition to decompose the heat load data to obtain IMF component signals; construct a dataset with the IMF component signals and meteorological data; Construct a heat load forecasting model, use the dataset to train the heat load forecasting model, and optimize the key parameters of the heat load forecasting model using the sparrow search algorithm during the training process; use the trained heat load forecasting model to predict the heat load of the measured integrated energy system.

[0007] Preferably, the grey wolf optimization algorithm is used to perform parameter optimization on the penalty factor and the number of mode decompositions of variational mode decomposition.

[0008] Preferably, the key parameters include learning factors, the number of neurons in the hidden layer, and the training period.

[0009] Preferably, the process of parameter optimization for the penalty factor and the number of mode decompositions is as follows: Step a. Initialize the parameters of the grey wolf optimization algorithm; the position of the grey wolf corresponds to the penalty factor and the number of mode decompositions; Step b. Take the position of the grey wolf as the optimization variable and the minimum envelope entropy as the fitness function; Step c. Select wolves from the grey wolf population according to the fitness of the grey wolf and based on wolves, and The position of the wolf updates the position of the gray wolf individual; Step d. Obtain the fitness of each gray wolf and update the optimal solution according to the fitness; Step e. Repeat Step c to Step d until the termination condition is reached, and output the penalty factor and modal decomposition number corresponding to the optimal solution.

[0010] Preferably, the method for optimizing the key parameters of the heat load prediction model is as follows: Step a. Initialize the sparrow population and construct a fitness function based on the fitting degree between the prediction result and the true value of the heat load prediction model; Step b. Divide the sparrow population into producers and followers according to the fitness of the sparrows; Step c. Update the position of the producers and update the position of the followers based on the position of the producers; Step d. Set a vigilant individual in the sparrow population; when the vigilant individual gives an early warning, update the position of the sparrow population; Step e. Obtain the fitness of each sparrow and update the optimal solution according to the fitness; Step f. Repeat Step b to Step e until the iterative termination condition is met, and output the key parameters corresponding to the optimal solution.

[0011] Preferably, the heat load prediction model adopts a long short-term memory model.

[0012] Preferably, the variational mode decomposition decomposes the original signal through the alternating direction method of multipliers and the iterative update strategy.

[0013] In a second aspect, the present invention provides a heat load prediction system based on multi-stage fusion optimization, which is used to execute the above-mentioned heat load prediction method based on multi-stage fusion optimization; the heat load prediction system includes a data acquisition module, a data processing module, and a heat load prediction module; the data acquisition module is used to acquire heat load data and meteorological data; the data processing module is used to decompose the heat load data to obtain IMF component signals; the heat load prediction module is used to perform heat load prediction according to the IMF component signals and meteorological data.

[0014] In a third aspect, the present invention provides a computer device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, the memory stores the computer program; the processor executes the above-mentioned heat load prediction method.

[0015] In a fourth aspect, the present invention provides a readable storage medium, storing a computer program; when the computer program is executed by a processor, it is used to implement the above-mentioned heat load prediction method.

[0016] The beneficial effects of the present invention are as follows: 1. The present invention uses the Grey Wolf Optimization Algorithm to replace the traditional empirical method to optimize the key parameters of variational mode decomposition, significantly improving the decomposition quality of variational mode decomposition. At the same time, the Sparrow Search Algorithm is used to globally optimize the hyperparameters of Long Short-Term Memory (such as the number of hidden layer nodes, learning rate, training epochs, etc.), overcoming the problems of low efficiency and easy entrapment in local optima of the traditional grid search, avoiding the blindness and computational complexity of the traditional grid search method, and improving the prediction accuracy and efficiency of the model.

[0017] 2. The present invention effectively combines the advantages of the variational mode decomposition algorithm in signal decomposition and the advantages of the Long Short-Term Memory model in time series prediction, providing a new idea for non-stationary heat load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is the overall flowchart of the present invention.

[0019] Figure 2 is the Grey Wolf hierarchy diagram of the present invention.

[0020] Figure 3 is the flowchart of the Grey Wolf Optimization Algorithm of the present invention.

[0021] Figure 4 is the schematic diagram of each IMF component obtained by using the variational mode decomposition method of the present invention.

[0022] Figure 5 is the structural schematic diagram of the Long Short-Term Memory model of the present invention.

[0023] Figure 6 is the flowchart of the Sparrow Search Algorithm of the present invention.

[0024] Figure 7 is the comparison diagram of prediction errors of different models.

[0025] Figure 8 is the comparison diagram of the predicted values and the true values of different models. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The present invention will be further described below with reference to the accompanying drawings.

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] Such as Figure 1As shown, a heat load prediction method based on multi-stage fusion optimization includes the following steps: Step 1: Obtain the historical heat load data of the tested integrated energy system, with a data sampling frequency of once every 1 hour. At the same time, collect meteorological data of the corresponding period in the area, including wind speed, wind direction, precipitation, air pressure, surface reflectivity, solar zenith angle, relative humidity, total radiation, direct radiation, dew point temperature, cloud cover, temperature and diffuse horizontal radiation.

[0029] Step 2: Figure 2 and Figure 3 As shown in the figure, the gray wolf optimization algorithm (GWO) is used to optimize the penalty factor α and the modal decomposition number K of the variational mode decomposition (VMD). The gray wolf optimization algorithm simulates the social hierarchy and hunting behavior of wolves. The wolf pack is divided into four levels: α The wolf, as leader, is responsible for decision making; β Wolves assist in decision making and serve as alternate leaders, helping α Wolves lead and hunt; δ Wolves perform reconnaissance and surveillance missions; ω The wolves are responsible for implementation. This hierarchical structure enables efficient collaboration through top-down information transfer: α The wolf determines the search direction. β Wolves optimize local areas, δ Wolf provides additional information, ω The wolf performs the specific search. The update process using the gray wolf optimization algorithm is as follows: 2-1. Initialize the parameters of the GWO algorithm, including the number of wolf packs, the number of iterations, and wolf packs at different levels; randomly generate the position of the gray wolf in the wolf pack; the position of the gray wolf corresponds to the penalty factor α and the modal decomposition number K of VMD.

[0030] 2-2. Take the position of the gray wolf as the optimization variable, minimize the envelope entropy as the fitness function, and use the GWO algorithm for optimization.

[0031] 2-3. The formula for setting the wolf pack to surround the prey is: In the formula, D is the distance between the gray wolf and the target prey in the current iteration; is the current iteration number; is the position vector of the prey; is the position vector of the gray wolf; and is the synergy coefficient vector, and its calculation formula is: in, is the convergence factor; and is an interval in the random vector.

[0032] During the iteration process, the convergence factor is in the interval linearly decreases. Select the 3 gray wolves with the best fitness in the wolf pack as wolves respectively, and guide the grass-roots wolves to surround the prey. This process can be expressed as: where are respectively the distances between the wolves and the target prey; are respectively affected by the wolves after ω the wolf vector positions; are respectively the coefficient vectors corresponding to the wolves.

[0033] According to the update rules of the GWO algorithm, update the positions and velocities of the gray wolf individuals, and calculate the fitness values of each gray wolf. Update the optimal solution according to the fitness values.

[0034] 2 - 4. Determine whether the termination condition is reached. If not, return to the above steps; if so, output the penalty factor α and the modal decomposition number K corresponding to the optimal solution.

[0035] Step 3: According to the penalty factor α and the modal decomposition number K corresponding to the optimal solution obtained in Step 2, use the VMD algorithm to decompose the heat load data. VMD is essentially a variational problem. Its core idea is to decompose complex non-linear signals into a series of intrinsic mode functions (IMFs) with specific center frequencies and bandwidths by constructing and solving a constrained variational model. The VMD algorithm transforms the signal decomposition problem into an optimization problem by introducing constraint conditions, so as to effectively process non-linear and non-stationary signals. As a typical non-linear time series, the heat load data usually contains multiple frequency components and complex dynamic change laws. Therefore, the VMD algorithm is very suitable for the decomposition and processing of heat load data. The advantage of the VMD algorithm lies in its solid mathematical theory foundation, which can adaptively determine the center frequency and bandwidth of each IMF, thus avoiding the modal aliasing problem commonly found in traditional signal decomposition methods (such as empirical mode decomposition EMD). The specific process of the VMD algorithm is as follows: The VMD algorithm decomposes the original signal into multiple IMF components through the alternating direction method of multipliers (ADMM) and the iterative update strategy Each IMF component is a frequency-modulated - amplitude-modulated signal (FM-AM signal) with adjustable amplitude and frequency, and its mathematical expression can be represented as: where, is the instantaneous amplitude function; is the phase function.

[0036] The phase function is non-decreasing, that is ≥0. The analytical signal of the mode function is calculated using the Hilbert transform to obtain the single-sided spectrum. The decomposed signal is mixed with its central frequency, and each mode spectrum is converted to the fundamental frequency. According to the square norm of the gradient direction of the modulation signal, the bandwidth of each mode is estimated, and its expression is: where, is the set of k mode components obtained by decomposition; is the set of the central frequencies of each component; * is the convolution symbol; is the Dirac function; is the original signal; K is the number of mode decompositions.

[0037] The quadratic penalty factor and the Lagrange multiplier operator are introduced to transform the constrained variational problem into an unconstrained variational problem, and the K intrinsic mode functions are updated, and its expression is:

[0038] where, is the Lagrange augmented function of; represents the inner product symbol.

[0039] If the convergence condition is satisfied, stop the iteration and output at least one IMF component signal, otherwise continue the iteration. The VMD decomposition diagram optimized by GWO is as shown in Figure 4 . According to the obtained IMF component signals and meteorological data, a data set is constructed, and the data set is divided into a training set and a test set in a ratio of 7:3.

[0040] Step Four, as shown in Figure 5As shown, a heat load prediction model is constructed; the heat load prediction model adopts an LSTM (Long Short-Term Memory) model; LSTM consists of three parts: a forget gate, an input gate, and an output gate, which can "remember" the required information and "forget" the unnecessary information. LSTM has strong robustness and fault tolerance capabilities and can fully approximate complex non-linear relationships. During the operation of LSTM, the input of the forget gate is the current input , and the output of the previous moment . These inputs are linearly combined and transformed into values between 0 and 1 through the Sigmoid activation function . The closer the output value is to 1, the more information is retained in the memory block. The closer the output value is to 0, the less information is retained. Its calculation formula is: In the formula, represents the output value of the forget gate; represents the output value of the previous moment; represents the current input value; represents the bias matrix of the forget gate; , represents the weight coefficient matrix.

[0041] The input gate includes a Sigmoid layer that determines the value to be updated and a tanh layer that creates a new candidate value vector. The Sigmoid layer of the input gate also linearly combines the current input and the output of the previous moment , and then compresses this linear combination through the Sigmoid activation function . The input gate is similar to the forget gate in function, but its function is to determine to what extent the input information at the current moment is added to the memory information flow. The corresponding formula is: In the formula, represents the output value of the input gate; represents the bias matrix of the Sigmoid layer of the input gate; represents the corresponding weight coefficient matrix of the Sigmoid layer of the input gate.

[0042] The tanh layer of the input gate linearly combines the input value and the output of the previous moment , and generates an alternative value of the memory cell state through the tanh function. Then, the alternative value is linearly combined with the memory cell state of the previous moment to obtain the updated memory cell state . The corresponding calculation formula is as follows: In the formula, is the state of the memory cell at the current moment; represents the bias matrix of the input gate tanh layer; is the weight coefficient matrix corresponding to the input gate tanh layer.

[0043] The output gate controls the information to be output in the memory cell through the Sigmoid activation function to obtain the weight coefficient , and then the output gate and the new cell state passing through the tanh layer to obtain the output value In the formula, represents the weight matrix; represents the bias matrix of the output gate; represents the output value of the output layer.

[0044] Step Five. As Figure 6 shown, use the Sparrow Search Algorithm (SSA) to optimize the learning factor, the number of neurons in the hidden layer, and the training cycle of the LSTM model. Compared with traditional methods (such as manual tuning) that rely on experience, have low efficiency and are prone to missing the optimal solution, LSTM is a complex deep learning model, and there may be complex non-linear relationships between its hyperparameters. SSA can better handle this non-linear relationship. In addition, the heat load data usually has the characteristics of non-linearity and non-stationarity and is affected by various factors such as weather, season, and time. The LSTM model optimized by SSA can better adapt to the non-linear and non-stationary characteristics of the heat load data and improve the robustness of the model. SSA is based on the behavior of sparrows and is a meta-heuristic intelligent optimization algorithm with strong global search and adaptive capabilities. Compared with traditional algorithms, it is more excellent in dealing with high-dimensional optimization problems and can quickly find excellent hyperparameter combinations to avoid local optima. Its core is to simulate the foraging and anti-predation of sparrows, and divide the population into three categories: producers, followers, and vigilant ones. Producers search for food widely with strong adaptability, followers snatch food, and vigilant ones move randomly without danger and guide others to escape when in danger, thereby balancing global and local searches to effectively find the optimal solution. The specific process of the Sparrow Search Algorithm is as follows: 5-1. Initial population Initially set the sparrow population, where the position of each sparrow is randomly determined. The population contains d×n sparrows, which is the candidate solution set for the optimization problem; set the initial dimension d , d is the number of optimization targets; n is the size of the population. The expression of the number matrix about sparrows is as follows: Among them, is the n th d -dimensional parameter of the

[0045] The fitness value of each sparrow is randomly assigned. As the iteration process progresses, the algorithm gradually updates the fitness value of each sparrow, thereby guiding the algorithm to gradually approach a better solution. The initial fitness value of the sparrow The expression is as follows: Among them, is the initial fitness value of different dimensions.

[0046] The training set is input into the LSTM model batch by batch, and the fitness is constructed according to the fitting degree between the prediction result of the LSTM model and the true value , and its calculation formula is as follows: 5-2. Sparrows are divided into producers and followers according to their fitness. The task of the producer is to continuously search for potential food resources. The equation for updating the position of the producer is as follows: In the formula, represents the t th i -dimensional parameter of the j th sparrow in the th iteration; represents the maximum number of iterations; ST is a uniform random number in (0, 1]; AL is the safety threshold, taking a uniform random number in [0, 1]; Q is the alarm value, taking [0.5, 1]; L is 1×d 's identity matrix.

[0047] The process of updating the position of the follower is as follows: If the follower can find sufficient food independently and does not rely on the producer, in order to maintain the balance of different roles in the population, this part of the follower will replace the position of some producers to complete the position update. If the producer finds food, the follower may preempt the position of the producer, thus triggering the position update process. The equation for updating the position of the follower is as follows: In the formula, Indicates the worst position occupied by the sparrow; Indicates the best position occupied by the sparrow; ; A is a 1×d matrix where the elements are randomly taken as 1 or -1.

[0048] If , it means that the current sparrow position is not good and no food has been caught, so the search range needs to be expanded. If , the sparrow position is better, but more efficient search and more accurate solution localization can still be achieved.

[0049] 5-3. Updating the position of the vigilant Set the number of vigilant individuals to one-fifth of the total population, and their initial positions are randomly generated. When the vigilant individuals give early warnings, the sparrow population is updated; according to the rules of iterative update of the positions of producers and followers in the sparrow population, the corresponding position update equations are: In the formula, represents the current optimal position of the population; is a control parameter subject to the standard normal distribution; K is a random number between [-1,1]; is an infinitesimal to prevent the denominator from being 0; represents the i th sparrow's fitness; represent the best fitness and the worst fitness of the sparrows respectively.

[0050] When > , it means that the current sparrow is at the edge of the population and is extremely vulnerable to attack. When , it indicates that the sparrow is aware of the danger and starts to move towards other sparrows to reduce the risk of being preyed upon.

[0051] 5-4. Obtain the fitness of each sparrow and update the optimal solution according to the fitness; 5-5. When the number of iterations is exhausted, the parameters stop updating, and the learning factor, the number of hidden layer neurons, and the training cycle corresponding to the optimal solution are output, so as to obtain the optimized heat load prediction model.

[0052] Step 6: Use the test set to evaluate the optimized heat load prediction model. During the evaluation process, the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Median Absolute Percentage Error (MAPE) are used as evaluation indicators, which are expressed as follows: In the formula, They are the Root Mean Square Error, Mean Absolute Error, and Median Absolute Percentage Error respectively; is the true value of the heat load; is the predicted value of the heat load; is the number of samples in the test set.

[0053] RMSE, MAE, and MAPE are commonly used indicators to measure the accuracy of the prediction model. Among them, the Root Mean Square Error represents the square root of the mean of the squares of the differences between the prediction results and the actual observed values. The Mean Absolute Error represents the average of the absolute values of the prediction errors. The Median Absolute Percentage Error represents the average of the percentage errors of each predicted value relative to the actual value. The smaller the values of the above indicators, the better the fitting effect of the prediction model, the higher the prediction accuracy, and the smaller the relative error.

[0054] Use the present invention and existing models (LSTM benchmark model, VMD-LSTM model, VMD-SSA-LSTM model) to predict the heat load sequence respectively; the LSTM benchmark model directly uses the LSTM model to predict the original heat load sequence; the VMD-LSTM model performs VMD decomposition on the original sequence and then inputs them into LSTM respectively; the VMD-SSA-LSTM model introduces SSA after VMD decomposition, optimizes the LSTM model input after VMD decomposition, and then predicts by LSTM; the prediction results of different heat load prediction methods for heat load prediction are shown in Table 1.

[0055] Table 1 Comparison of error indicators of different experimental models Method RMSE(KW) ⬇ MAE(KW) ⬇ MAPE(%) ⬇ LSTM 322.2657 277.0474 10.8264 VMD - LSTM 301.1182 259.4235 10.0618 VMD - SSA - LSTM 267.6962 217.4442 8.4085 The present invention 220.4453 175.2300 6.7161 Figure 7 is the comparison chart of the prediction errors of the above various prediction methods. It can be seen from the figure that the present invention has a smaller error performance compared with other models. Figure 8This is a comparison chart of the prediction results of the present invention and the above other models on the heat load dataset. Further explanation: Aiming at the problems of complex parameter coupling and low convergence rate in multi-algorithm collaborative modeling, the present invention designs a phased optimization and nested training process, realizing a double improvement in model stability and prediction accuracy, highlighting the potential of this model for more accurate and reliable prediction, which is of great significance for the practical applications of energy management and thermal system optimization.

[0056] To verify the necessity and non-obviousness of each technical module in the present invention, ablation experiments were further carried out. The core components (VMD, SSA, GWO) of this model were gradually removed, and the degree of model performance degradation was observed. After removing the GWO optimization, the VMD decomposition parameters were set empirically. Compared with the present invention, the prediction error increased by about 1.6924%, indicating that the GWO parameter optimization has a significant improvement effect on the VMD decomposition quality; after removing the SSA, the LSTM model showed overfitting, and the error increased by about 3.3457% compared with the present invention, indicating that the SSA is indispensable for noise reduction and sequence smoothness improvement; when directly using the LSTM for modeling, the prediction accuracy decreased most significantly, and the error almost doubled compared with the present invention, indicating that the non-stationarity of the original sequence seriously affects the model performance. The above results fully show that: the collaborative optimization process constructed by the deep coupling of algorithms in the present invention is not a simple superposition of existing technologies, but generates a non-linear collaborative enhancement effect in the heat load prediction task, significantly superior to each sub-component model, and has outstanding technological progressiveness.

Claims

1. A thermal load prediction method based on multi-stage fusion optimization, characterized in that: It includes the following steps: Obtain the historical heat load data and meteorological data of the integrated energy system to be measured; Optimize the penalty factor and the number of mode decompositions of variational mode decomposition, and decompose the heat load data using variational mode decomposition according to the optimal penalty factor and the number of mode decompositions to obtain IMF component signals; construct a data set with the IMF component signals and meteorological data; Construct a heat load prediction model, use the data set to train the heat load prediction model, and optimize the key parameters of the heat load prediction model using the sparrow search algorithm during the training process; Use the trained heat load prediction model to predict the heat load of the integrated energy system to be measured.

2. The thermal load prediction method based on multi-stage fusion optimization according to claim 1, wherein: Use the grey wolf optimization algorithm to optimize the penalty factor and the number of mode decompositions of variational mode decomposition.

3. The method for predicting heat load based on multi-stage fusion optimization according to claim 1, wherein: The key parameters mentioned above include the learning factor, the number of neurons in the hidden layer, and the training period.

4. A thermal load prediction method based on multi-stage fusion optimization according to claim 1, characterized in that: The process of optimizing the penalty factor and the number of mode decompositions is as follows: Step a. Initialize the parameters of the grey wolf optimization algorithm; the position of the grey wolf corresponds to the penalty factor and the number of mode decompositions; Step b. Take the position of the grey wolf as the optimization variable and the minimum envelope entropy as the fitness function; Step c. Select wolves from the gray wolf population according to the fitness of the gray wolves, and update the positions of the gray wolf individuals based on the positions of the wolves; Step d. Obtain the fitness of each grey wolf and update the optimal solution according to the fitness; Step e. Repeat Step c to Step d until the termination condition is reached, and output the penalty factor and the number of mode decompositions corresponding to the optimal solution.

5. A heat load prediction method based on multi-stage fusion optimization according to claim 1, characterized in that: The method for optimizing the key parameters of the heat load prediction model is as follows: Step a. Initialize the sparrow population and construct a fitness function based on the fitting degree between the prediction result and the true value of the heat load prediction model; Step b. Divide the sparrow population into producers and followers according to the fitness of the sparrows; Step c. Update the position of the producers and update the position of the followers based on the position of the producers; Step d. Set warning individuals in the sparrow population; if a warning individual gives a warning, update the position of the sparrow population; Step e. Obtain the fitness of each sparrow and update the optimal solution according to the fitness; Step f. Repeat Step b to Step e until the iteration termination condition is met, and output the key parameters corresponding to the optimal solution.

6. The thermal load prediction method based on multi-stage fusion optimization according to claim 1, wherein: The heat load prediction model mentioned above adopts a long short-term memory model.

7. A thermal load prediction method based on multi-stage fusion optimization according to claim 1, characterized in that: The variational mode decomposition mentioned above decomposes the original signal through the alternating direction multiplier method and the iterative update strategy.

8. A heat load prediction system based on multi-stage fusion optimization, characterized in that: It is used to execute a heat load prediction method based on multi-stage fusion optimization described in Claim 1; this heat load prediction system includes a data acquisition module, a data processing module, and a heat load prediction module; The data acquisition module is used to collect heat load data and meteorological data; the data processing module is used to decompose the heat load data to obtain IMF component signals; the heat load prediction module is used to perform heat load prediction based on the IMF component signals and meteorological data.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: The memory stores a computer program; the processor executes the heat load prediction method described in any one of Claims 1-7.

10. A readable storage medium stores a computer program; characterized in that: When the computer program is executed by the processor, it is used to implement the heat load prediction method described in any one of Claims 1-7.

Images