Air conditioning system load prediction method and device, medium and equipment

Abstract

The invention discloses an air conditioning system load prediction method and device, a medium and equipment, and relates to the technical field of load prediction. The method comprises the following steps: firstly, acquiring historical load data of the air conditioning system and load influence factors of the historical load data in the same period, then constructing a preliminary load prediction model based on a seasonal difference autoregression moving average model according to the historical load data, and determining a prediction residual error of the preliminary load prediction model to the historical load data; then training the generalized regression neural network to obtain a residual error prediction model according to the prediction residual error of the load influence factors and the initial load prediction model on the historical load data, and finally inputting the load data to be predicted into the initial load prediction model to carry out initial load prediction, and correcting the load data through the residual error prediction model to obtain an initial load prediction result. And determining the predicted load of the air conditioning system at the target moment according to the initial load and the correction. According to the invention, the convergence speed, the prediction efficiency and the prediction accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of load forecasting, and in particular to a method, device, medium and equipment for forecasting load of an air-conditioning system. Background Art

[0002] At present, the energy consumption of air conditioning systems in buildings accounts for about 40% to 50% of the total building energy consumption, so there is a great potential for energy saving. Accurately predicting the cooling load of air conditioning systems not only helps to reduce energy consumption and improve the efficiency of air conditioning systems, but also provides theoretical and practical support for the control of air conditioning systems.

[0003] In the prior art, air conditioning load prediction methods are generally divided into physical modeling methods and data-driven methods. The physical modeling method calculates the air conditioning load through mechanism modeling, using detailed building parameters, equipment operation data and personnel flow conditions. However, due to the characteristics of air conditioning systems such as nonlinearity, strong coupling, and large lag, it is difficult to establish a universal and accurate air conditioning load model. In recent years, with the development of the computer industry, especially the advancement of artificial intelligence technology, data-driven methods including support vector machines, neural network technology and other machine learning algorithms are widely used in air conditioning load prediction. The data-driven method starts from the characteristics of the data itself, considers its changing trend, and obtains the predicted value by training the model with a large amount of historical data and related variable data. It has the advantages of low cost, high flexibility, high accuracy and strong generalization ability.

[0004] However, when traditional machine learning algorithms are applied to air conditioning load prediction, the accuracy of air conditioning load prediction is often low when the network structure is simple, and when the accuracy of air conditioning load prediction is to be improved, a complex network structure is usually used, which leads to slow convergence speed and low prediction efficiency. In other words, the traditional air conditioning load prediction method cannot take into account both the prediction efficiency and prediction accuracy of air conditioning load. Summary of the invention

[0005] Based on this, it is necessary to provide an air conditioning system load prediction method, device, medium and equipment to address the above technical problems.

[0006] The present invention adopts the following technical solutions: The present invention provides a load prediction method for an air-conditioning system. The method first obtains historical load data of the air-conditioning system and load influencing factors of the same period as the historical load data. Then, a preliminary load prediction model is constructed based on a seasonal difference autoregressive sliding average model according to the historical load data, and the prediction residual of the preliminary load prediction model for the historical load data is determined. Then, a generalized regression neural network is trained according to the load influencing factors and the prediction residual of the preliminary load prediction model for the historical load data to obtain a residual prediction model. Finally, the load data to be predicted can be input into the preliminary load prediction model for preliminary load prediction. The residual prediction model is used to determine the residual corresponding to the preliminary prediction according to the load influencing factors to be predicted so as to correct it, thereby determining the predicted load of the air-conditioning system at a target time according to the preliminary load and the correction.

[0007] The present invention provides an air conditioning system load prediction device, comprising: An acquisition module, used to acquire historical load data of the air-conditioning system and load influencing factors during the same period of the historical load data; A preliminary training module is used to train the seasonal difference autoregressive moving average model based on historical load data to obtain a preliminary load forecasting model; determine the prediction residual of the preliminary load forecasting model for the historical load data; The residual training module is used to train the generalized regression neural network according to the load influencing factors and the prediction residuals of the preliminary load forecasting model for the historical load data to obtain the residual forecasting model; A preliminary prediction module is used to input the load data to be predicted into the preliminary load prediction model, extract the time series characteristics in the load data to be predicted through the preliminary load prediction model, and determine the preliminary load of the air-conditioning system at the target time according to the time series characteristics; The residual prediction module inputs the load influencing factors to be predicted into the residual prediction model, extracts the relationship characteristics between the load influencing factors and the air-conditioning load through the residual prediction model, and determines the correction of the preliminary load based on the relationship characteristics; the predicted load of the air-conditioning system at the target time is determined based on the sum of the preliminary load and the correction to the preliminary load.

[0008] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned air conditioning system load prediction method is implemented.

[0009] The present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned air conditioning system load prediction method when executing the program.

[0010] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects: The present invention splits and processes the relevant data of the load prediction of the air-conditioning system. On the one hand, the seasonal difference autoregressive sliding average model is used to capture the periodic characteristics of the load data of the air-conditioning system, so as to fit the change law of the load data of the air-conditioning system, make a preliminary prediction on the load data of the air-conditioning system to be predicted, and determine the preliminary load prediction result of the air-conditioning system at the target time under the change law; on the other hand, the generalized regression neural network is used to capture the nonlinear relationship between the load influencing factors and the load ignored in the first aspect, so as to further correct the preliminary load prediction result by using the nonlinear approximation ability of the generalized regression neural network, and by decoupling the entire air-conditioning system load prediction process, the complexity of the network structure is reduced, the convergence speed and prediction efficiency are improved, and the prediction accuracy is improved at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 A schematic flow chart of a method for predicting load of an air conditioning system provided by the present invention; Figure 2 A schematic diagram of a SARIMA model prediction process provided by the present invention; Figure 3 A schematic diagram of the GRNN model structure provided by the present invention; Figure 4 A schematic diagram of an NRBO-GRNN model optimization process provided by the present invention; Figure 5 A schematic diagram of a load prediction device for an air conditioning system provided by the present invention; Figure 6 A schematic diagram of a computer device for implementing a method for predicting load of an air conditioning system provided by the present invention. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0013] In order to solve the problems existing in the existing methods, this patent proposes a prediction model that combines the Seasonal Autoregressive Integrated Moving Average (SARIMA) model with the Generalized Regression Neural Network (GRNN) and the Newton-Raphson-based optimizer (NRBO). The SARIMA model is used for preliminary prediction and the GRNN network is used for secondary learning to further improve the prediction accuracy and handle complex nonlinear problems. The input parameters and parameter adjustment process required for the prediction model are simplified, and the prediction performance is more outstanding. At the same time, it has significant advantages in optimization accuracy, convergence speed and stability.

[0014] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0015] Figure 1 The present invention is a flow chart of a method for predicting load of an air conditioning system, which specifically includes the following steps: S101: Obtain historical load data of the air conditioning system and load influencing factors during the same period of the historical load data.

[0016] S102: training a seasonal difference autoregressive moving average model based on historical load data to obtain a preliminary load forecasting model; and determining a prediction residual of the preliminary load forecasting model for the historical load data.

[0017] S103: According to the load influencing factors and the prediction residuals of the preliminary load forecasting model for the historical load data, a generalized regression neural network is trained to obtain a residual prediction model.

[0018] S104: inputting the load data to be predicted into the preliminary load prediction model, extracting the time series characteristics in the load data to be predicted through the preliminary load prediction model, and determining the preliminary load of the air-conditioning system at the target time according to the time series characteristics.

[0019] S105: Input the load influencing factors to be predicted into the residual prediction model, extract the relationship characteristics between the load influencing factors and the air-conditioning load through the residual prediction model, and determine the corrected residual of the preliminary load according to the relationship characteristics; determine the predicted load of the air-conditioning system at the target time according to the sum of the preliminary load and the corrected residual.

[0020] For the convenience of explanation, the following description is only based on the server as the execution subject. The server mentioned in the present invention can be a server set up on the business platform, or a device such as a desktop computer, a notebook computer, etc. that can execute the solution of the present invention.

[0021] In general, data-driven methods for predicting the load of air-conditioning systems mostly use historical data of air-conditioning loads, which have obvious periodicity on a time scale of years or days. Therefore, in one or more embodiments of the present invention, the server of the business platform can obtain historical load data of the air-conditioning system. At the same time, the present invention uses a seasonal difference autoregressive moving average model for load prediction of air-conditioning systems. The SARIMA model is a time series model that comprehensively considers the dependency, seasonality and linear trend of the time series but does not consider the nonlinear trend. If this model is used to fit the data, the information contained in the nonlinear trend in the time series data will be omitted. The GRNN model has a strong nonlinear approximation ability and can extract the information contained in the nonlinear trend. Combining the SARIMA model and the GRNN model to establish a SARIMA-GRNN combined model, the characteristics of the information provided by the two prediction models are comprehensively utilized, which can not only make up for the defect that the SARIMA model cannot fit the nonlinear part, but also improve the fitting accuracy of the overall model to make the prediction more accurate, thereby improving the stability of the air-conditioning system.

[0022] Among them, the SARIMA model is used to make preliminary predictions using historical load data. SARIMA can describe time series that show periodic characteristics due to seasonal changes (changes in years, months, weeks, and days), and when dealing with time series problems, it can simultaneously consider the correlation between data and the interference of random factors. It has high prediction accuracy and is often used to fit time series problems. In one or more embodiments of the present invention, the historical load data of the air-conditioning system is used as the input of the SARIMA model to predict the load of the air-conditioning system at historical moments, and the prediction residual of the predicted air-conditioning system load is calculated based on the historical load data at historical moments, that is, the difference between the actual load value and the predicted value.

[0023] The establishment of the SARIMA model includes the following three steps: (1) Use the unit root test (Augmented Dickey-Fuller, ADF) test to stabilize the time series of the original load history data and determine the difference order (d, D); (2) Use the autocorrelation function (Auto-Correlation Function, ACF) diagram and the partial autocorrelation coefficient (Partial Auto-Correlation Function, PACF) diagram to determine the possible model parameters (p: number of non-seasonal autoregressive terms; q: number of non-seasonal moving average terms; P: number of seasonal autoregressive terms; Q: number of seasonal moving average terms; s: length of seasonal cycle); (3) Use the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) to screen the optimal model to obtain the optimal model. Figure 2 As shown, Figure 2 It is a schematic diagram of the SARIMA model prediction process in the present invention.

[0024] Specifically, the server can use a unit root test to determine the stationarity of historical load data. When the historical load data has a trend, the historical load data is differentiated. When the historical load data has seasonality, the historical load data is seasonally differentiated until the historical load data is stable. The non-seasonal difference order d and seasonal difference order D are determined when the seasonal difference autoregressive sliding average model is used to predict the air-conditioning system load.

[0025] According to the autocorrelation function graph and partial autocorrelation coefficient graph corresponding to the historical load data, the number of non-seasonal autoregressive items to be selected when using the seasonal difference autoregressive moving average model for air conditioning system load forecasting is determined. p , non-seasonal moving average items q , seasonal autoregressive terms P , seasonal moving average items Q and the length of the seasonal cycle s .

[0026] According to the number of candidate non-seasonal autoregressive items, the number of non-seasonal moving average items, the number of seasonal autoregressive items, the number of seasonal moving average items and the length of the seasonal cycle, a plurality of candidate seasonal difference autoregressive moving average models are formed, and the candidate seasonal difference autoregressive moving average models are screened by the Akaike information criterion and the Bayesian information criterion, and the optimal seasonal difference autoregressive moving average model is used as the preliminary load forecasting model.

[0027] The formula for the SARIMA model is as follows: ,in, , , , , .

[0028] In the formula, x ( t ) is the air conditioning system load time series, is the stable historical load data after difference, B is the lag operator, is the difference operator, F ( B s ) f ( B ) is a seasonal autoregressive model, f ( B )for p Order autoregressive polynomial, f 1. f 2. … f p is the non-seasonal autoregressive parameter, F ( B s ) is the seasonal autoregressive polynomial, F 1. F 2. … F P for P Order seasonal autoregressive parameters, I ( B s ) i ( B ) is the seasonal moving average model, i ( B )for q Order moving average polynomial, i 1. i 2. … i q is the seasonal moving average polynomial, I ( B s ) is the non-seasonal moving average polynomial, I 1. I 2. … I Q for Q The seasonal moving average parameter, e ( t ) is Gaussian noise.

[0029] Furthermore, when performing air conditioning load prediction, the selection of prediction model input data has an important impact on the prediction speed of the model and the accuracy of the final output prediction result. In one or more embodiments of the present invention, in addition to obtaining the historical load data of the air conditioning system, the server can also obtain the load influencing factors of the same period as the historical load data, where the load influencing factors may include one or more combinations of outdoor temperature, indoor temperature, solar radiation, relative humidity, holiday attributes and population density.

[0030] Since the load influencing factors are diverse and their influence on the air-conditioning load varies, the factors with greater influence can be found among the many load influencing factors and determined as input variables.

[0031] Therefore, the server can collect various load influencing factors of the historical load data of the air-conditioning system at the same period according to the set time interval within the set historical data collection cycle, and use Pearson correlation analysis to determine the load influencing factors that have a greater impact on the load of the air-conditioning system, and use them as sample data to train the GRNN model. And the prediction residual corresponding to the prediction of the historical load data of the air-conditioning system by the SARIMA model is used as the label of the GRNN model, and the sample data is input into the generalized regression neural network to obtain the predicted prediction residual, and the generalized regression neural network is trained for multiple rounds with the minimum deviation between the predicted prediction residual and the label as the optimization goal. Here, the output of the GRNN model can be denormalized to obtain the predicted value of the GRNN model for the prediction residual. In order to avoid neuron saturation and eliminate the dimensional differences between variables, the sample data needs to be normalized to accelerate the optimization process and improve the training accuracy. In addition, the total sample data collected can be divided into training set and test set in a ratio of 3:1 to ensure the accuracy of the GRNN model during training and better generalization ability when applied.

[0032] The GRNN neural network model uses the SARIMA model to perform secondary learning on the prediction residuals of historical load data. By learning the relationship between the prediction residuals of the SARIMA model and various load influencing factors, the prediction results of the SARIMA model can be further corrected. That is, the GRNN model uses the residual error to compensate the SARIMA model, and can capture the nonlinear relationship that the SARIMA model fails to accurately predict. The GRNN model uses the corrected residual of the predicted load as output, and the final prediction value = the preliminary load predicted by the SARIMA model + the corrected residual of the GRNN model. In this way, using the GRNN model to compensate for the residual of the SARIMA model can effectively improve the accuracy of load forecasting.

[0033] The GRNN model is a neural network model based on nonlinear regression theory. It is a type of radial basis function network (RBF). The GRNN consists of four layers: input layer, pattern layer, summation layer, and output layer. Figure 3 As shown, Figure 3 This is a schematic diagram of the GRNN model structure in the present invention.

[0034] Figure 3 The input sample is The output sample is The number of neurons in the input layer is equal to the dimension of the input variable, which can directly transmit the input to the pattern layer; each neuron in the pattern layer corresponds to a different sample, and the number of neurons corresponds to the number of training samples. n The transfer function of a neuron , i = 1, 2, …, I .

[0035] In the formula, X is the input sample of the model, X i For the i The observed value corresponding to each neuron is s is the smoothing factor.

[0036] The summation layer includes two types of neurons, one of which performs arithmetic summation of the outputs of neurons in all pattern layers, and its transfer function is .

[0037] Another type of neuron performs weighted summation of the outputs of all pattern layers, with the weight being n The dependent variable (output vector) of training samples (input samples) Y n No. j elements, and its transfer function is: .

[0038] In the formula, y nj For the model layer n neurons and the summation layer j The summed neuron connection weights.

[0039] Output layer, the number of neurons is equal to the dimension of the output vector J , each neuron divides the output of the summation layer, j The output of the neuron corresponds to the prediction result j Elements: ,j =1, 2, …, J .

[0040] In the formula, y i The output layer j The output of a neuron.

[0041] The GRNN neural network relies on training samples, and only one parameter, the smoothing factor, needs to be adjusted in model construction. s When is very large, the prediction results are close to the dependent variable (output vector) of all training samples (input samples) Y The network has a strong generalization ability. s When it approaches 0, the prediction result is close to the dependent variable (output vector) of the training sample (input sample) closest to the prediction point. Y , the network approximation effect is good, but if it encounters a prediction point that is not included in the training sample, the prediction effect will be very poor and overfitting will occur. In order to obtain a suitable smoothing factor s In one or more embodiments of the present invention, the Newton-Raphson optimization algorithm is used to optimize the parameters. The NRBO algorithm is an optimization algorithm that combines the Newton-Raphson method and the heuristic search technique. The algorithm aims to solve complex optimization problems through efficient local search capabilities and global search capabilities.

[0042] The NRBO algorithm uses the Newton-Raphson method to find the search area to define the search path, and uses the Newton-Raphson Search Rule (NRSR) and the Trap Avoidance Operator (TAO) to explore the search domain. Figure 4 As shown, Figure 4 It is a schematic diagram of the optimization process of the NRBO-GRNN model in the present invention.

[0043] Specifically, in one or more embodiments of the present invention, the server may construct a generalized regression neural network corresponding to each individual in each round of iteration, and determine the fitness value corresponding to the individual according to the prediction accuracy of the prediction residual of the constructed generalized regression neural network for the preset test sample; and determine the worst population position and the best population position of the population according to the fitness value of the individual. According to the worst population position and the best population position, the Newton-Raphson optimization algorithm is used to update the individuals in the population, and through multiple rounds of iterations, the global optimal solution in the population after the iteration is used as the optimal smoothing factor.

[0044] As the iteration proceeds, the adaptive parameters dBalance exploration and development, more exploration in the early stage, and concentrated development in the later stage. Repeat the search and update process, gradually optimize the smoothing factor until the maximum number of iterations is reached. The final output global optimal solution Xb is the optimized optimal smoothing factor, which is used to improve the prediction performance of GRNN.

[0045] Specifically, first, a random population can be generated by the following formula: .

[0046] in, xnj Indicates n The individual j The parameters to be optimized, lb , ub are the upper and lower bounds of the parameters to be optimized, respectively. rand is a random number between (0,1); N p is the number of individuals in the population; dim is the dimension of the parameter to be optimized. In the present invention, the parameter to be optimized is the smoothing factor. dim is 1.

[0047] The population matrix is ​​as follows: .

[0048] In the formula, For the population, Each row in corresponds to an individual, and each column corresponds to a dimension of the parameter to be optimized.

[0049] The Newton-Raphson search rule is as follows: , , , .

[0050] In the formula, is the vector generated in the update step of the Newton-Raphson search rule, The current iteration n The solution corresponding to each individual is xnj middle j When 1 is taken, , randn is a normally distributed random number with mean 0 and variance 1, X w is the worst population position of the current iteration, X b is the best population position of the current iteration, is the process vector generated when generating new individuals, which means exploring the worse areas in the search space based on the worse individuals in the population. is the process vector generated when generating new individuals, which represents the search for further developing better areas based on better individuals in the population, is the amount of change when generating a new individual, is a random number between (0,1), Search parameters for Newton-Raphson.

[0051] Computing NRSR can enhance the current solution by helping it update its position. This formula increases the search capability of NRBO through random parameters and better balances exploration and exploitation capabilities. In order to find the optimal solution in the search space and eventually converge to the global solution, a parameter called adaptive coefficient can be applied. d to enhance the algorithm. d The expression is:

[0052] In the formula, IT Indicates the current iteration number, Max_IT Indicates the maximum number of iterations.

[0053] In order to maintain a balance between the exploration and exploitation phases, the adaptive coefficient d Adaptation during the iteration process significantly reduces the number of iterations while taking into account the random behavior of the optimization process, increasing diversity and avoiding local optimality, thereby improving the NRBO algorithm. r To improve the development of the proposed NRBO, the optimization algorithm further refines the search near the already found high-quality solutions to obtain better local solutions, and focuses the search in a local area of ​​the search space to maximize the use of the information of the currently obtained best solution. This parameter guides the population in the right direction. r The expression is: .

[0054] In the formula, a and b is a random number between (0,1), r 1 and r 2 are unequal, distinct integers selected randomly from the population.

[0055] The search method proposed by the following formula is benign when searching locally, but has limitations when searching globally: .

[0056] In the formula, To update The new vector position obtained is is the individual position at the ITth iteration, indicating the smoothing factor value of the current individual, is the set of smoothing factor values ​​of all individuals in the entire population in the IT-th iteration, and are two individual positions randomly selected from the population.

[0057] The search strategy proposed by the following formula is benign in global search, but has limitations in local search: .

[0058] In the formula, For consideration X b , , The new vector position is used to further improve the search effect.

[0059] NRBO uses both of the above equations to improve the exploration and exploitation phases. The new position vector at the next iteration is as follows: , .

[0060] In the formula, is the solution vector at the IT+1th iteration, For consideration , and The generated new position vector is used for local exploration to balance global search and local search.

[0061] Using Trap Avoidance Operation (TAO) can significantly change By the following formula, when rand The value is less than DF When : , , .

[0062] In the formula, β is a binary number 1 or 0, is a uniform random number between (-1,1), is a uniform random number between (-0.5, 0.5), is the average value of the smoothing factors of all individuals in the IT-th iteration, It is a local optimal solution, a better smoothing factor value generated during the local exploration process.

[0063] Due to the parameters m 1 and m 2 The randomness of the selection makes the population more diverse, thus getting rid of the constraints of the local optimal solution.

[0064] The pseudo code of NRBO optimization algorithm is as follows: initialization: 1. Choose the population size ( N p ), maximum number of iterations ( Max_IT ) and the determining factor ( DF ) is the appropriate value.

[0065] 2. Initialize the random ensemble position.

[0066] 3. Evaluate initial fitness and specify X b (optimal population position) and X w (Worst population position).

[0067] Main loop: while ( IT < Max_IT ); oven =1: N p ; forge =1: dim ; In the interval [1, N p ] Pick unequal random integers r 1 , r 2 ; Determine the solution ; end for ; ifrand < DF ; Finding a solution ; ; else ; Determine the solution ; end if ; renew X b and X w ; end for ; IT = IT +1; end while ; Termination: If the termination condition is met, return the best solution X b .

[0068] The implementation steps of SARIMA model construction and NRBO algorithm optimization of GRNN neural network are as follows Figure 5 As shown, Figure 5 The following is a schematic diagram of a SARIMA-NRBO-GRNN model prediction process in the present invention, including: 1. Data preparation and preprocessing. Collect and clean the historical load data of the air-conditioning system and the time series data corresponding to the load influencing factors of the same period of the historical load data, handle missing values ​​and outliers, normalize the data, and further divide the sample data set.

[0069] For example, first collect data and pre-process it. Within a set historical data collection period (e.g., no less than one year), collect the historical load data of the air-conditioning system and various load influencing factors of the same period at a set time interval (e.g., 1 hour), and use Pearson correlation analysis to determine the load influencing factors that have a greater impact on the air-conditioning system load. I The total sample data collected can be divided into training set and test set in a ratio of 3:1 to ensure the accuracy of the model during training and better generalization during testing.

[0070] Here we can use the linear function normalization method: .

[0071] In the formula, is the normalized data, is the data to be normalized, is the minimum value of the data type to be normalized, The maximum value of the data type to be normalized.

[0072] At the same time, in order to make the obtained data reflect the true value of the original dimension, the output prediction results need to be denormalized.

[0073] 2. Use ADF test to make the original time series stable and determine the difference order ( d , D ).

[0074] 3. Determine possible SARIMA model parameters through the ACF graph and the PACF graph. p , q , P , Q , s ).

[0075] 4. The optimal SARIMA model was obtained by using Akaike information criterion AIC and Bayesian information criterion BIC to screen the optimal model.

[0076] 5. Input the historical load data of the air conditioning system into the SARIMA model to predict the air conditioning system load at the historical target time and calculate the prediction residual. Prediction residual = actual value of the air conditioning system load at the historical target time - SARIMA prediction value.

[0077] 6. Initialize the GRNN model parameters, including determining the input and output variables of the GRNN model, initializing the smoothing factor, etc. Input variables may include: outdoor temperature at the historical moment corresponding to the historical load data, outdoor temperature at the moment before the historical moment, indoor temperature at the historical moment, load forecast residual, solar radiation at the historical moment, relative humidity at the historical moment, holiday attributes at the historical moment, and occupancy density in the building at the historical moment.

[0078] 7. Initialize NRBO algorithm parameters, including population size, maximum number of iterations, and determining factors.

[0079] 8. For each GRNN model corresponding to each population individual, use the GRNN to make predictions and calculate the root mean square error RMSE As the fitness value of the individual: .

[0080] in, y i is the prediction residual predicted by the GRNN model, is the annotation for GRNN model training, i.e. the actual prediction residual. M is the number of iterations.

[0081] 9. Start iteration and apply the Newton-Raphson search rule to each individual in the population so that the fitness value gradually decreases.

[0082] 10. Introduce the trap avoidance operation (TAO) to adjust the search path of individuals in the population to avoid falling into the local optimum and update the parameters of GRNN.

[0083] 11. Select individuals with lower fitness to form a new population, re-evaluate the fitness values ​​of the updated population individuals, and determine the global optimal position and global optimal fitness value.

[0084] 12. Determine whether to end the optimization process based on the preset termination conditions: maximum number of iterations or fitness threshold. Take the smoothing factor corresponding to the lowest fitness value generated during the iteration as the optimal smoothing factor.

[0085] 13. The SARIMA-NRBO-GRNN prediction model is constructed using the optimal smoothing factor, and the test samples are predicted. The prediction correction value is output and summed with the preliminary prediction value of the SARIMA model to obtain the final prediction value.

[0086] Through the above process, the SARIMA model and the GRNN model are successively constructed and trained, so as to obtain a preliminary load forecasting model and a residual forecasting model, so that the load data to be predicted is input into the preliminary load forecasting model, the time series characteristics in the load data to be predicted are extracted through the preliminary load forecasting model, and the preliminary load of the air-conditioning system at the target time is determined according to the time series characteristics, and the load influencing factors to be predicted are input into the residual prediction model, and the relationship characteristics between the load influencing factors and the air-conditioning load are extracted through the residual prediction model, and the corrected residual of the preliminary load is determined according to the relationship characteristics, and the predicted load of the air-conditioning system at the target time is determined according to the sum of the preliminary load and the corrected residual.

[0087] based on Figure 1 The load forecasting method for the air-conditioning system shown in the figure first obtains the historical load data of the air-conditioning system and the load influencing factors of the same period as the historical load data, then builds a preliminary load forecasting model based on the seasonal difference autoregressive sliding average model according to the historical load data, and determines the prediction residual of the preliminary load forecasting model for the historical load data, and then trains the generalized regression neural network according to the load influencing factors and the prediction residual of the preliminary load forecasting model for the historical load data to obtain the residual prediction model, and finally the load data to be predicted can be input into the preliminary load forecasting model for preliminary load prediction, and the residual prediction model is used to determine the residual corresponding to the preliminary prediction according to the load influencing factors to be predicted to correct it, so as to determine the predicted load of the air-conditioning system at the target time according to the preliminary load and the correction.

[0088] The present invention splits and processes the relevant data of the load prediction of the air-conditioning system. On the one hand, the seasonal difference autoregressive sliding average model is used to capture the periodic characteristics of the load data of the air-conditioning system, so as to fit the change law of the load data of the air-conditioning system, make a preliminary prediction on the load data of the air-conditioning system to be predicted, and determine the preliminary load prediction result of the air-conditioning system at the target time under the change law; on the other hand, the generalized regression neural network is used to capture the nonlinear relationship between the load influencing factors and the load ignored in the first aspect, so as to further correct the preliminary load prediction result by using the nonlinear approximation ability of the generalized regression neural network, and by decoupling the entire air-conditioning system load prediction process, the complexity of the network structure is reduced, the convergence speed and prediction efficiency are improved, and the prediction accuracy is improved at the same time.

[0089] In building energy consumption, the air conditioning system is one of the main energy-consuming equipment. Using the historical load of the air conditioning system to predict the air conditioning load is a commonly used data-driven method. Because the air conditioning load has obvious periodicity on a time scale of years or days, the present invention considers using the SARIMA model for load prediction of the air conditioning system. And combining the NRBO algorithm to optimize the GRNN neural network for air conditioning system load prediction can achieve accurate energy consumption management. Combined with real-time data, the model can predict short-term and long-term air conditioning loads. More accurate load predictions can help optimize the operation strategy of the air conditioning system, reduce unnecessary energy consumption, and achieve energy-saving effects. In a home environment, using the optimized GRNN model for air conditioning load prediction can adjust the air conditioning operation according to the activity patterns of family members and weather changes. Accurate load prediction enables the air conditioning system to respond to load changes more quickly and improve user experience and comfort.

[0090] The load of the air-conditioning system is cyclical, and its prediction involves a variety of complex nonlinear relationships and random fluctuations. The SARIMA model can capture and process the cyclical fluctuations in the time series, but the prediction of complex nonlinear relationships may not be accurate enough. GRNN is good at dealing with complex nonlinear relationships and random fluctuations, and the NRBO algorithm can further optimize its performance. The NRBO algorithm uses the fast convergence characteristics of the Newton-Raphson method to quickly find accurate solutions in local searches, improve its prediction performance and generalization ability, and find optimized GRNN parameters in fewer iterations, thereby improving training efficiency. By combining heuristic search algorithms, NRBO can effectively search globally and avoid falling into local optimality. The trap avoidance operation is introduced to increase the diversity of solutions by adjusting the search path, thereby improving the robustness and stability of the prediction model.

[0091] After using the SARIMA model for preliminary prediction, the prediction residuals are calculated. The GRNN neural network model can use these residuals for secondary learning to correct and improve the preliminary prediction. This error correction mechanism can significantly improve the accuracy of the prediction. The GRNN neural network model compensates for the prediction results of the SARIMA model by learning the relationship between the residual error and other influencing factors in the SARIMA model prediction, and can capture the nonlinear relationship that the SARIMA model fails to accurately predict. This model structure of preliminary prediction and secondary learning can significantly improve the accuracy and reliability of air-conditioning load prediction. The combined model is more stable when facing different types of data changes. When the seasonality and trend changes of the data are obvious, the SARIMA model provides stable predictions; when the data has complex nonlinear changes, the GRNN model effectively compensates. For example, the air-conditioning load in summer and winter has obvious periodic changes, and the SARIMA model can provide a better baseline prediction. The GRNN model provides more accurate corrections through secondary learning and compensation for the nonlinear effects of weather changes, user behavior, etc. on the air-conditioning load. This complementary mechanism enables the combined model to maintain high prediction accuracy in various scenarios, reduces the model's dependence on single data characteristics, and improves the robustness and stability of the prediction.

[0092] When applying the air conditioning system load prediction method provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0093] The above is an air conditioning system load prediction method provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding air conditioning system load prediction device, such as Figure 6 shown.

[0094] Figure 6 A schematic diagram of a load prediction device for an air conditioning system provided by the present invention includes: An acquisition module 201 is used to acquire historical load data of the air conditioning system and load influencing factors of the same period as the historical load data; The preliminary training module 202 is used to train the seasonal difference autoregressive moving average model according to the historical load data to obtain a preliminary load forecasting model; determine the prediction residual of the preliminary load forecasting model for the historical load data; The residual training module 203 is used to train the generalized regression neural network according to the load influencing factors and the prediction residuals of the preliminary load prediction model for the historical load data to obtain the residual prediction model; The preliminary prediction module 204 is used to input the load data to be predicted into the preliminary load prediction model, extract the time series characteristics in the load data to be predicted through the preliminary load prediction model, and determine the preliminary load of the air conditioning system at the target time according to the time series characteristics; The residual prediction module 205 inputs the load influencing factors to be predicted into the residual prediction model, extracts the relationship characteristics between the load influencing factors and the air-conditioning load through the residual prediction model, and determines the correction of the preliminary load according to the relationship characteristics; and determines the predicted load of the air-conditioning system at the target time according to the sum of the preliminary load and the correction to the preliminary load.

[0095] For the specific definition of the air conditioning system load prediction device, please refer to the definition of the air conditioning system load prediction method above, which will not be repeated here. Each module in the above-mentioned air conditioning system load prediction device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0096] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 The air conditioning system load prediction method is provided.

[0097] The present invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The air conditioning system load prediction method is provided.

[0098] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0099] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A method for predicting load of an air conditioning system, characterized in that: include: Obtain the historical load data of the air conditioning system and the load influencing factors during the same period of the historical load data; The seasonal difference autoregressive moving average model is trained according to the historical load data to obtain a preliminary load forecasting model; the prediction residual of the preliminary load forecasting model for the historical load data is determined; According to the load influencing factors and the prediction residuals of the preliminary load forecasting model for historical load data, the generalized regression neural network is trained to obtain the residual prediction model; Inputting the load data of the air-conditioning system to be predicted into the preliminary load prediction model, extracting the periodic time series characteristics in the load data to be predicted through the preliminary load prediction model, and determining the preliminary load of the air-conditioning system at the target time according to the time series characteristics; Input the load influencing factors to be predicted into the residual prediction model, extract the nonlinear relationship characteristics between the load influencing factors and the air-conditioning load through the residual prediction model, and determine the corrected residual of the preliminary load; The predicted load of the air conditioning system at the target time is determined based on the sum of the preliminary load and the corrected residual.

2. The air conditioning system load prediction method according to claim 1, characterized in that: The seasonal difference autoregressive moving average model is trained according to the historical load data to obtain a preliminary load forecasting model, which specifically includes: The unit root test is used to determine the stability of historical load data. When the historical load data has a trend, the historical load data is differentiated. When the historical load data has seasonality, the historical load data is seasonally differentiated until the historical load data is stable. The non-seasonal difference order and seasonal difference order are determined when the seasonal difference autoregressive sliding average model is used for air conditioning system load forecasting. According to the autocorrelation function graph and partial autocorrelation coefficient graph corresponding to the historical load data, the number of non-seasonal autoregressive items, the number of non-seasonal moving average items, the number of seasonal autoregressive items, the number of seasonal moving average items and the length of the seasonal cycle are determined when the seasonal difference autoregressive moving average model is used for air conditioning system load forecasting; According to the number of candidate non-seasonal autoregressive items, the number of non-seasonal moving average items, the number of seasonal autoregressive items, the number of seasonal moving average items and the length of the seasonal cycle, a plurality of candidate seasonal difference autoregressive moving average models are formed, and the candidate seasonal difference autoregressive moving average models are screened by the Akaike information criterion and the Bayesian information criterion, and the optimal seasonal difference autoregressive moving average model is used as the preliminary load forecasting model.

3. The air conditioning system load prediction method according to claim 1, characterized in that: The training of the generalized regression neural network based on the load influencing factors and the prediction residuals of the preliminary load forecasting model for the historical load data specifically includes: The load influencing factors of the same period as the historical load data are used as sample data, and the prediction residuals are used as annotations; The sample data is input into the generalized regression neural network to obtain the predicted prediction residual. The generalized regression neural network is trained for multiple rounds with the minimum deviation between the predicted prediction residual and the annotation as the optimization goal. In each round of training, the Newton-Raphson optimization algorithm is used to optimize the smoothing factor of the generalized regression neural network.

4. The air conditioning system load prediction method according to claim 3, characterized in that: The Newton-Raphson optimization algorithm is used to optimize the smoothing factor of the generalized regression neural network, specifically including: The smoothing factors of the generalized regression neural network are generated by random algorithms to form a population, and each individual in the population is a generated smoothing factor; Constructing a generalized regression neural network corresponding to each individual, and determining the fitness value corresponding to the individual according to the prediction accuracy of the prediction residual of the constructed generalized regression neural network for the preset test sample; determining the worst population position and the best population position of the population according to the fitness value of the individual; According to the worst population position and the best population position, the Newton-Raphson optimization algorithm is used to update the individuals in the population. After multiple rounds of iterations, the global optimal solution in the population after iteration is used as the optimal smoothing factor.

5. The air conditioning system load prediction method according to claim 1, characterized in that: The load influencing factors include: one or more combinations of outdoor temperature, indoor temperature, solar radiation, relative humidity, holiday attributes and population density.

6. The air conditioning system load prediction method according to claim 5, characterized in that: The load influencing factors of the same period as the historical load data are obtained include: During the same period of historical load data, obtain the corresponding data of various load influencing factors according to the preset time interval; According to the historical load data of the air-conditioning system and the corresponding data of each load influencing factor, the correlation between each load influencing factor and the historical load data of the air-conditioning system is determined through Pearson correlation analysis; According to the preset correlation threshold, various load influencing factors are screened.

7. An air conditioning system load prediction device, characterized in that: include: An acquisition module, used to acquire historical load data of the air conditioning system and load influencing factors of the same period as the historical load data; A preliminary training module is used to train the seasonal difference autoregressive moving average model based on historical load data to obtain a preliminary load forecasting model; determine the prediction residual of the preliminary load forecasting model for the historical load data; The residual training module is used to train the generalized regression neural network according to the load influencing factors and the prediction residuals of the preliminary load forecasting model for the historical load data to obtain the residual forecasting model; A preliminary prediction module is used to input the load data to be predicted into the preliminary load prediction model, extract the time series characteristics in the load data to be predicted through the preliminary load prediction model, and determine the preliminary load of the air-conditioning system at the target time according to the time series characteristics; The residual prediction module inputs the load influencing factors to be predicted into the residual prediction model, extracts the relationship characteristics between the load influencing factors and the air-conditioning load through the residual prediction model, and determines the correction of the preliminary load based on the relationship characteristics; the predicted load of the air-conditioning system at the target time is determined based on the sum of the preliminary load and the correction to the preliminary load.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images