Thermal load prediction method, system and equipment based on multivariate input and storage medium

Abstract

The invention discloses a thermal load prediction method, system and equipment based on multivariate input and a storage medium, relates to the technical field of heat supply control, and is used for solving the technical problem that multi-step prediction of thermal loads is realized by adopting multivariate time sequence data in the prior art. According to the technical scheme, the method mainly comprises the steps of obtaining multivariate time sequence data of the heat supply system; establishing a data set based on the multivariate time series data, and performing normalization processing on the data set to obtain a normalized data set; the normalized data set comprises a training set; a thermal load multi-step prediction model is constructed; the thermal load multi-step prediction model comprises a sequence-to-sequence model, an encoder, a decoder and a time sequence attention mechanism; the input data of the thermal load multi-step prediction model comprises a training set; and obtaining thermal load multi-step prediction data by using the thermal load multi-step prediction model. According to the method, the multi-step prediction of the heat supply system heat load data is realized by using the heat load multi-step prediction model.

Description

Technical Field

[0001] The invention relates to the technical field of heating control, and in particular to a method for predicting load of a heating system. Background Art

[0002] With the rapid development of smart heating systems, load forecasting has become a basic technology for optimizing the configuration of heating equipment in advance and achieving energy-saving operation. Heating load forecasting refers to the prediction of the dynamic overview of the heat load of the heating system at each moment in the short term in the future. Obtaining the heat load change curve in the future period is of great significance to supporting the optimization scheduling and refined control of the operation of the heating system and ensuring on-demand heating. However, the existing load forecasting technology still has shortcomings. In terms of data utilization, there is a lack of comprehensive utilization of multivariate time series; the heating system involves a variety of influencing factors, such as outdoor temperature, historical load and weather forecast, etc. Traditional methods are usually unable to fully integrate and mine the potential dynamic relationship of these data, and the prediction effect of heat load is limited, especially when the ambient temperature changes drastically or the user demand fluctuates greatly. It is difficult to effectively model the complex dynamic nonlinear characteristics in the heating load. In terms of data prediction, the prediction accuracy and adaptability are insufficient; when facing real-time load changes, the existing technology is often not ideal in prediction accuracy, and the generalization ability is limited, and it is difficult to adapt to the complex heating needs in different scenarios. Based on actual application needs, obtaining the heat load change curve in the future is of great significance to supporting the optimal scheduling and refined control of the heating system operation and ensuring that the heating system provides heating on demand. Summary of the invention

[0003] The invention solves the technical problem of realizing multi-step prediction of heat load by using multivariate time series data in the prior art.

[0004] The present invention proposes a heat load prediction method based on multi-input, the method is implemented based on the following system, the system includes a heating device and a heating system, and the method includes:

[0005] S1. Obtaining multivariate time series data of the heating system;

[0006] S2. establishing a data set based on the multivariate time series data, and normalizing the data set to obtain a normalized data set; the normalized data set includes a training set;

[0007] S3. A method for constructing a multi-step thermal load prediction model; the multi-step thermal load prediction model comprises a sequence-to-sequence model, an encoder, a decoder and a temporal attention mechanism; the input data of the multi-step thermal load prediction model comprises the training set;

[0008] S4. Obtaining heat load multi-step prediction data using the heat load multi-step prediction model.

[0009] Furthermore, the step of S1 includes:

[0010] S11. Constructing a directed connection graph based on the heating system;

[0011] S12. Constructing a steady-state hydraulic simulation model based on the directed connection graph; the steady-state hydraulic simulation model simulates the operating state of the heating system;

[0012] S13. Construct the multivariate time series data based on the operating status; the multivariate time series data includes target time series data and external cause time series data.

[0013] Furthermore, the step of S4 includes:

[0014] S41. Establishing a seq2seq sequence-to-sequence model; inputting the time series data into the seq2seq sequence-to-sequence model to obtain heat load data at all times within the prediction range;

[0015] S42. Inputting the heat load data at all times within the prediction range into the encoder, using the encoder to obtain the average value of the hidden state at the current time, and obtaining the weighted hidden state of the encoder;

[0016] S43. Using the decoder to obtain multi-step prediction data of heat load;

[0017] S44. The decoder introduces a temporal attention mechanism to obtain an updated weighted hidden state of the encoder; the temporal attention mechanism includes an attention weight and a weighted sum;

[0018] S45. Use the updated encoder weighted hidden state as the input of the decoder, and use the S43 method to obtain multi-step prediction data of thermal load.

[0019] Further, the weighted hidden state of the encoder in S42 is:

[0020]

[0021] In the formula, represents the weighted hidden state of the encoder, where D represents the number of neurons in the hidden layer of the encoder; t represents the time; τ represents the prediction range; h t =f e (h t-1 ,x t ), represents the hidden state at the current moment, where f e (·) is the LSTM neuron in the encoder; x is the external time series outdoor air temperature data; x t is the external time series at time t, h t-1 is the hidden state of the previous time step; Represents the hidden state of the encoder hidden layer at time t; W represents the length of the input training set and the external factor time series.

[0022] Furthermore, the step of S43 includes:

[0023] S431. Determine the hidden state of the decoder hidden layer, and when the hidden state of the decoder hidden layer is within the prediction range, predict the heat load;

[0024] S432. Weight the encoder hidden state Input to the decoder and compared with the decoder hidden state d W+i-1 Concatenate and get the concatenated vector:

[0025]

[0026] S433. The concatenated vector is input to the fully connected layer, and after linear transformation, the heat load multi-step prediction data is output as:

[0027]

[0028] Where y is the target time series heat load data; Represents the predicted value of the target time series; Represents the linear transformation weight vector of the fully connected layer; represents the state of the fully connected layer, and D' represents the neurons of the fully connected layer

[0029] number; Represents the bias vector of the fully connected layer; represents the bias of the fully connected layer.

[0030] Furthermore, the steps of S44 are as follows:

[0031] S441. Get attention weight; set the model memory unit to The hidden state at the previous moment is d t'-1 ; The hidden state h of the encoder hidden layer at time t t With the memory cell c' t-1 Concatenate and get the concatenated vector:

[0032]

[0033] Obtain the correlation between the hidden state of the encoder hidden layer at time t and the concatenated vector:

[0034]

[0035] In the formula, Represents the linear transformation weight vector of the temporal attention mechanism layer; is the input weight matrix, which represents the state of the temporal attention mechanism layer; is the state weight matrix, which represents the state of the temporal attention mechanism layer; Represents the bias vector of the temporal attention mechanism layer;

[0036] S442. Perform weighted summation on the attention weight and the weighted hidden state of the encoder to obtain an updated weighted hidden state of the encoder:

[0037]

[0038] In the formula, β t't Represents the attention weight at the current moment.

[0039] The present invention proposes a multi-step prediction device for heat load based on multi-input, the device is implemented based on the following system, the system includes a heating device and a heating system, and the device includes:

[0040] Data acquisition module: used to acquire multivariate time series data of the heating system;

[0041] Data preprocessing module: used to establish a data set based on the multivariate time series data, and also used to normalize the data set to obtain a normalized data set; the normalized data set includes a training set;

[0042] Model building module: a method for building a heat load multi-step prediction model; the heat load multi-step prediction model includes a seq2seq sequence to sequence model, an encoder, a decoder and a temporal attention mechanism; the input data of the heat load multi-step prediction model includes the training set;

[0043] Data prediction module: used to obtain heat load multi-step prediction data using the heat load multi-step prediction model.

[0044] The present invention also proposes a control device for a multi-step prediction method of thermal load based on multi-inputs, the device comprising a processor and a memory, wherein the memory is used to store a computer program, and when the microprocessor reads the computer program on the memory, it executes the multi-step prediction method of thermal load based on multi-inputs.

[0045] The present invention further proposes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for predicting heat load based on multi-input proposed by the present invention is implemented.

[0046] The present invention also proposes a computer program product, comprising a computer program / instruction, which implements the steps of the multi-step prediction method of heat load based on multi-element input when executed by a processor.

[0047] Beneficial technical effects of the present invention:

[0048] First, the present invention obtains the multivariate time series data of the heating system by configuring the steady-state hydraulic simulation model constructed by the heating system; through the dynamic modeling of the multivariate time series data, the accuracy of the heat load prediction of the heating system is significantly improved, and the reliability of the hydraulic simulation model is enhanced, which provides beneficial technical support for the efficient operation and optimization of the heating system.

[0049] Second, the present invention constructs a thermal load prediction model; currently, the heating Internet of Things stores a large number of historical time series such as outdoor temperature and heat load. The present invention effectively models the dynamic spatiotemporal patterns in multivariate time series based on recurrent neural networks, realizes accurate prediction of various target time series, and constructs a thermal load prediction model to obtain the change curve of the thermal load. Through the constructed multi-step prediction model of thermal load, the multi-step prediction data of thermal load is obtained by using the model, and the accurate prediction of the thermal load of the heating system is realized. The model comprehensively utilizes diversified input data for comprehensive prediction of the thermal load change curve; the model combines the Seq2Seq sequence to sequence framework, encoder and decoder, and can efficiently and collaboratively process multivariate time series data and capture the time series characteristics in the input data; the model supports the prediction of multi-step load changes, provides the load change trend in the future time period, and provides a scientific basis for the optimal scheduling of the heating system; the model introduces the time series attention mechanism, which significantly improves the ability to dynamically allocate weights to key time steps and exogenous variables.

[0050] On the whole, the present invention provides higher heat load prediction accuracy and model adaptability, provides a prediction data basis for the adaptive regulation of water conservancy simulation models, effectively supports the optimal control effect of the heating system, and promotes the intelligent development of the heating system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a structural diagram of the multi-step prediction model of heat load in implementation mode 1;

[0052] Figure 2 A schematic diagram of the steps of obtaining heat load multi-step prediction data using a heat load multi-step prediction model in implementation mode 1;

[0053] Figure 3 It is a schematic diagram of constructing a directed connection graph based on a heating system in implementation mode 2;

[0054] Figure 4 It is a schematic diagram of constructing a steady-state hydraulic simulation model based on a directed connection graph in implementation mode 2;

[0055] Figure 5A flowchart of the recursive prediction model of the LSTM network in implementation mode 3 for obtaining a heat load prediction value based on the outdoor air temperature sequence and heat load time series data;

[0056] Figure 6 A flow chart of heat load prediction results obtained by using a recursive prediction model based on a long short-term memory network based on the outdoor air temperature time series and heat load time series data in implementation mode 3;

[0057] Figure 7 It is a structural diagram of the i-th direct prediction model of the LSTM network in implementation mode three;

[0058] Figure 8 This is a prediction flow chart of the long short-term memory network direct prediction model based on the LSTM network in implementation mode three;

[0059] Fig. 9 It is a structural diagram of the LSTM-seq2seq prediction model in implementation mode three;

[0060] Fig.10 It is a structural diagram of the TALSTM-seq2seq prediction model in implementation mode 3;

[0061] Fig.11 It is a graph of the hourly heat load and the hourly outdoor air temperature change in the fourth implementation mode;

[0062] Fig.12 This is a graph of selected hourly heat load changes output by each prediction model in the fourth implementation mode. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0064] Embodiment 1: A heat load prediction method based on multi-input is described in this embodiment. The method is implemented based on the following system, the system includes a heating device and a heating system, and the method includes:

[0065] S1. Obtaining multivariate time series data of the heating system;

[0066] S2. establishing a data set based on the multivariate time series data, and normalizing the data set to obtain a normalized data set; the normalized data set includes a training set;

[0067] S3. Method for constructing a multi-step prediction model for heat load (e.g. Figure 1 As shown); the heat load multi-step prediction model includes a sequence-to-sequence model, an encoder, a decoder and a temporal attention mechanism; the input data of the heat load multi-step prediction model includes the training set;

[0068] S4. Using the heat load multi-step prediction model to obtain heat load multi-step prediction data (such as Figure 2 shown).

[0069] In this embodiment, by acquiring multivariate time series data of the heating system and processing the data, a multi-step prediction model for heat load is constructed, and the multi-step prediction data for heat load is calculated through the multi-step prediction model for heat load, thus forming a heat load prediction method based on multivariate input as a whole. This method has significant technical effects and practical value in engineering applications in the field of heating control; by acquiring multivariate time series data and normalizing it, the data distribution differences are effectively eliminated, and the reliability of data analysis and the generalization ability of the model are improved; the normalized data set, especially the optimization of the training set, makes the constructed multi-step prediction model for heat load have higher accuracy and stability; the multi-step prediction ability of the model significantly enhances the prediction of future demand for the heating system, and can provide accurate data support for the hydraulic simulation of the heating system and the operation of the heating equipment network.

[0070] Embodiment 2: This embodiment is a further example of S1 in the heat load prediction method based on multi-input described in Embodiment 1. The steps of S1 include:

[0071] S11. Construct a directed connection graph based on the heating system (such as Figure 3 shown);

[0072] S12. Construct a steady-state hydraulic simulation model based on the directed connection graph (such as Figure 4 The steady-state hydraulic simulation model simulates the operating state of the heating system;

[0073] S13. Construct the multivariate time series data based on the operating status; the multivariate time series data includes target time series data and external cause time series data.

[0074] This embodiment further defines a method for obtaining multivariate time series data of the heating system. In this embodiment, the heat network characteristic identification aims to achieve a comprehensive analysis, evaluation and prediction of the operating status of the heating system under different control strategies, and provide rich and complete training samples for the pipeline leakage fault diagnosis model. According to graph theory, the heating system is abstracted as a directed connection graph composed of pipes and nodes. Hot water in the heating system is transported as full pipe flow, and pressure changes will be transmitted in the hot water at the speed of sound. Compared with the cycle of hydraulic regulation, the response time of hydraulic parameters is very short. Therefore, the dynamic hydraulic conditions can be regarded as a series of continuous quasi-steady-state hydraulic conditions. A steady-state hydraulic simulation model of the heating system is established using the node flow balance equation and the loop pressure balance equation. In view of the uncertainty of pipeline impedance parameters in the heating network, an impedance identification method based on genetic algorithm is adopted to verify the pipeline impedance parameters in the hydraulic simulation model, so as to further improve the accuracy of hydraulic simulation.

[0075] Embodiment 3: This embodiment is a further example of S4 in the heat load prediction method based on multi-input described in Embodiment 1, and the steps of S4 include:

[0076] S41. Establishing a seq2seq sequence-to-sequence model; inputting the time series data into the seq2seq sequence-to-sequence model to obtain heat load data at all times within the prediction range;

[0077] S42. Inputting the heat load data at all times within the prediction range into the encoder, using the encoder to obtain the average value of the hidden state at the current time, and obtaining the weighted hidden state of the encoder;

[0078] S43. Using the decoder to obtain multi-step prediction data of heat load;

[0079] S44. The decoder introduces a temporal attention mechanism to obtain an updated weighted hidden state of the encoder; the temporal attention mechanism includes an attention weight and a weighted sum;

[0080] S45. Use the updated encoder weighted hidden state as the input of the decoder, and use the S43 method to obtain multi-step prediction data of thermal load.

[0081] In this embodiment, the method of using a decoder to calculate multi-step prediction data of heat load includes a method of determining the hidden state of a hidden layer of a decoder, using the hidden state of the decoder to predict the heat load, and concatenating it with the hidden state of the decoder, and then using a fully connected layer linear transformation to output the multi-step prediction data of heat load.

[0082] First, the selection of prediction methods for building a multi-step heat load prediction model. There are three commonly used methods for short-term heat load multi-step prediction: recursive prediction (Recursive Strategy), direct prediction (Direct Strategy) and multi-input multi-output (Multi-Input Multi-Output, also known as sequence to sequence, Sequence to Sequence, abbreviated as seq2seq) prediction.

[0083] The recursive prediction method is to first establish a single-step prediction model f Recursive (·) to predict the heat load at time t+1, and then input the predicted value as a prediction variable into the established single-step prediction model to predict the heat load at time t+2. This process is iterative until the predicted values ​​of the heat load at all times within the prediction range τ are obtained. The recursive prediction method can be described by the following formula.

[0084]

[0085] In the formula, represents the predicted value of heat load at time t+i; x t represents the value of the external time series (outdoor air temperature) at time t; y t represents the value of the target time series (heat load) at time t; W represents the length of the target and exogenous time series used as model input.

[0086] The recursive prediction method only needs to train a single-step prediction model, and the modeling process is simple and the computational overhead is low. However, since the predicted value at the previous moment will be used to predict the heat load at the subsequent moment, the prediction error at the previous moment will affect the prediction accuracy at the subsequent moment, that is, the error accumulation phenomenon will occur. In addition, the training goal of the recursive prediction is not consistent with the goal of the multi-step prediction of the heat load. The training goal of the single-step prediction model is to minimize the error between the predicted value and the actual value of the heat load at the next moment, while the goal of the multi-step prediction of the heat load is to minimize the error between the predicted value and the actual value of the heat load within the prediction range.

[0087] The direct prediction method is to subdivide the multi-step prediction of heat load into several independent prediction tasks, that is, to establish τ independent prediction models to predict the heat load value at each moment within the prediction range. The mathematical description of this method is shown in the following formula.

[0088]

[0089] In the formula, f Direct,i (·) represents the i-th direct prediction model.

[0090] The direct prediction method does not have the problem of error accumulation because it establishes τ independent prediction models, and the model training objective is consistent with the multi-step prediction objective of heat load. The prediction variable of the direct prediction method does not need to use the outdoor air temperature at a future moment, so the prediction accuracy is not affected by the prediction accuracy of the outdoor air temperature. However, the direct prediction method assumes that the heat load values ​​within the prediction range are independent of each other, ignoring the temporal correlation between them, and because it needs to train multiple prediction models, the computational overhead will also increase with the increase of the prediction range.

[0091] The multi-input multi-output prediction method aims to train a seq2seq prediction model f seq2seq (·) to simultaneously output the heat load values ​​at all times within the prediction range. The mathematical description of this method is shown in the following formula.

[0092]

[0093] The multi-input multi-output prediction method only needs to build a prediction model, and the computational overhead is much less than the direct prediction method. Moreover, the training goal of the prediction model is consistent with the multi-step prediction goal of the heat load, and in theory, the optimal prediction result can be obtained.

[0094] This implementation further limits the method for setting the time series data of the heat load multi-step prediction model. There are many variables commonly used in heat load prediction, but only a few key parameters such as time, outdoor air temperature, and historical heat load need to be selected as prediction variables to obtain more accurate heat load prediction results. Considering that heating companies currently rarely conduct continuous monitoring of user behavior, the only meteorological parameter collected is generally the outdoor air temperature. Therefore, only two time series, outdoor air temperature and heat load, are selected for multi-step prediction of heat load. Among them, the heat load time series is called the target time series, and the outdoor air temperature time series is called the external cause time series. Use vector Represents an external time series of duration M, using the vector Represents a target time series with a duration of M. The element x in the vector i and i The multi-step heat load forecasting task can be described as: at time t, using the target and external time series of a certain length to predict the target time series (y t+1 ,y t+2 ,...,y t+τ ), the variable τ is also called the prediction horizon.

[0095] Then, the construction method of the load forecasting model and the method of calculating the multi-step forecasting data of the thermal load by the model are further defined.

[0096] (1) Recursive prediction model based on LSTM network

[0097] The recursive prediction method uses LSTM network to establish a prediction model. The prediction model consists of an input layer, a hidden layer consisting of two layers of LSTM neurons, and an output layer. The model input is the target and external factor time series with a duration of W. is the input of the prediction model update process at time t, the vector and Respectively represent the hidden states of the first and second layer LSTM neurons at time t, and the variable D 1 and D 2 represents the number of LSTM neurons in the first and second layers, It represents the predicted value of the target time series at time W+1 by the prediction model. The detailed update process of the recursive prediction model based on the LSTM network is as follows:

[0098] At time t, the first layer of LSTM neurons uses the vector [x t ;y t ] and the hidden state of the previous moment The hidden state at the current moment is calculated according to the following formula: And pass it to the second layer of LSTM neurons;

[0099] The second layer of LSTM neurons uses the hidden state of the previous moment and the hidden state transmitted by the first layer of LSTM neurons The hidden state at the current moment is calculated by the following formula

[0100] At time W, as shown in the following formula, the output layer undergoes a linear transformation according to the hidden state of the second layer LSTM neurons Get the predicted value of the target time series

[0101]

[0102] In the formula, f LSTM (·) represents LSTM neurons; Represents the linear transformation weight vector of the output layer, which is a learnable parameter.

[0103] Figure 5 The structure of the recursive prediction model based on the LSTM network is shown in Figure 6 The prediction process of the recursive prediction model based on the LSTM network is shown in . In the prediction process of the recursive prediction model based on the LSTM network, the prediction value at the previous moment will be used for the prediction at the subsequent moment. If the prediction value of all heat loads within the prediction range τ is to be obtained, the prediction process needs to be called iteratively.

[0104] (2) Direct prediction model based on LSTM network

[0105] The direct prediction method also uses LSTM network to establish a prediction model. In order to obtain the heat load within the prediction range τ, the direct prediction method establishes τ prediction models based on LSTM network. The input variables of each prediction model are exactly the same, which are the target and external time series with a duration of W, and the output variable is the heat load at the target moment to be predicted. The structure of the prediction model, the format of the input variables, and the update process of the hidden layer are exactly the same as those of the recursive prediction model based on LSTM network. The only difference is the output layer. For the i-th direct prediction model based on LSTM network, the update process of its output layer can be described by the following formula. The following figure introduces the prediction process of the direct prediction model based on LSTM network.

[0106]

[0107] Figure 7 The structure of the i-th direct prediction model based on the LSTM network is shown in Figure 8 The prediction process of the direct prediction model based on LSTM network is shown in .

[0108] (3) LSTM-seq2seq prediction model

[0109] The LSTM-seq2seq prediction model uses an encoder-decoder architecture, using two LSTM networks to perform encoding and decoding respectively. represents the hidden state of the LSTM encoder hidden layer at time t, the variable D represents the number of neurons in the encoder hidden layer, and the vector represents the weighted hidden state of the encoder, the vector represents the hidden state of the hidden layer of the LSTM decoder at time t′, and the variable D′ represents the number of neurons in the hidden layer of the decoder.

[0110] Fig. 9 The structure of the LSTM-seq2seq prediction model shown in .

[0111] The detailed updating process of the LSTM encoder is as follows: (1) At time t, the LSTM neuron uses the external time series x t and the hidden state h of the previous time step t-1 , the hidden state h at the current moment is calculated by the following formula t ; (2) Calculate the average value of the hidden state of the LSTM encoder at each time according to the following formula to obtain the weighted hidden state of the encoder: Then the encoder weighted hidden state Used to update the hidden state of the LSTM decoder at each moment.

[0112] h t =f e (h t-1 ,x t ),

[0113]

[0114] In the formula, f e (·) represents the LSTM neurons in the encoder.

[0115] The detailed update process of the LSTM decoder is as follows: First, when the time step t' is not greater than W, the LSTM decoder concatenates the observation value y of the target time series t' and the encoder weighted hidden state Then use the concatenation vector and the hidden state d at the previous moment t'-1 , update the hidden state d at the current moment by the following formula t' ; Then, when the time step t' is greater than W, the LSTM decoder concatenates the predicted values ​​of the target time series and the encoder weighted hidden state Then update the hidden state d at the current moment according to the following formula t' ; Finally, when predicting the heat load within the prediction range, the weighted hidden state of the splicing encoder is as shown in the following formula: and the LSTM decoder hidden state d W+i-1 , the concatenated vector The input is sent to a fully connected layer containing D' neurons, which then outputs the predicted value of the heat load through a linear transformation.

[0116]

[0117] In the formula, f d (·) represents the LSTM neurons in the decoder; Represents the linear transformation weight vector of the fully connected layer, a learnable parameter; Represents the state-input weight matrix of the fully connected layer, a learnable parameter; Represents the bias vector of the fully connected layer, a learnable parameter; Represents the bias of the fully connected layer, a learnable parameter.

[0118] (4) TALSTM-seq2seq prediction model

[0119] The prediction range of the heat load multi-step prediction task is generally large, and accordingly, the time series used for model input is also relatively long. Although the LSTM network can alleviate the long-range dependency problem in time series modeling, when the input time series is very long, the LSTM-seq2seq prediction model may also lose the information of the input sequence. Construct a TALSTM-seq2seq prediction model and use the temporal attention mechanism to adaptively learn the dynamic temporal patterns in the input time series. The structure of the TALSTM-seq2seq prediction model is as follows: Fig.10 Compared with the LSTM-seq2seq prediction model, the TALSTM-seq2seq prediction model only adds a temporal attention mechanism layer to adaptively select the hidden state of the LSTM encoder at each moment, and replace the average value of the LSTM encoder hidden state with the weighted sum of the attention weight and the LSTM encoder hidden state as the final encoder weighted hidden state.

[0120] This implementation method constructs a multi-step prediction model for heat load through the above process, and the implementation results of the steps show that it is advisable to use the multi-step prediction model for heat load to predict the short-term heat load profile of the heating system. In the multi-step prediction model for heat load, this implementation method further identifies the time series characteristics of the heating system by judging the hidden state of the decoder hidden layer, and realizes the effective use of time series data; using the hidden state to predict the heat load and splicing it with the hidden state helps to integrate the prediction information characteristics; combining the fully connected layer and the linear transformation, the continuity and stability of the multi-step prediction output can effectively control the accuracy. This implementation method enhances the accuracy and generalization ability of heat load prediction, and further provides technical support for the regulation and optimization of the heating system.

[0121] Embodiment 4: This embodiment is a further example of a heat load prediction method based on multi-input described in embodiment 1 through experiments. The experiments are analysis of actual application of load prediction, including:

[0122] (1) Project Dataset

[0123] The hourly heat load and hourly outdoor air temperature time series of an actual heating system are used as the data set for case analysis. The data set contains a total of 7223 samples from two heating seasons from November 1, 2014 to March 31, 2015 and from November 1, 2015 to March 31, 2016. Fig.11 The following is an excerpt of the hourly heat load and hourly outdoor air temperature change curves, specifically an excerpt of the hourly heat load and hourly outdoor air temperature change curves for a period of 14 days. According to the multi-step heat load prediction method introduced above, a corresponding prediction model is established to predict the heat load change curve of the heating system in the next day, that is, the prediction range τ = 24.

[0124] The samples from December 17, 2015 to February 14, 2016 were selected to form the test set, and the remaining samples were selected to form the training set. Then, the minimum-maximum normalization method was used to preprocess the training set and the test set, and the original hourly heat load and hourly outdoor temperature time series were linearly mapped to the [0,1] value range.

[0125] After data preprocessing, four types of prediction models were established using the training set, and then the performance of each prediction model was evaluated using the test set.

[0126] (2) Model hyperparameter setting and model training

[0127] 1) Parameter settings

[0128] For the recursive and direct prediction models based on the LSTM network, the duration of the input hourly heat load and hourly outdoor air temperature time series is set to 48. For the LSTM-seq2seq prediction model and the TALSTM-seq2seq prediction model, the duration of the input hourly heat load time series is set to 48, and the duration of the input hourly outdoor air temperature time series is set to 72.

[0129] For all prediction models, the learning rate of the Adam optimization algorithm is set to 0.001, the batch size K of the mini-batch gradient descent algorithm is 128, and the maximum number of training rounds is 50.

[0130] For the recursive and direct prediction models based on LSTM networks, the number of LSTM neurons in the two layers of their hidden layers is set to be the same. Similarly, the hidden layers of the encoder and decoder in the LSTM-seq2seq prediction model and the TALSTM-seq2seq prediction model are set to have the same number of neurons. The candidate number of hidden layer neurons in all prediction models is set to {16, 32, 64, 128}. The optimal number of hidden layer neurons is determined by the grid search algorithm. The following table shows the setting of the number of hidden layer neurons in each prediction model.

[0131] Table 1 Setting of the number of hidden layer neurons in each prediction model

[0132]

[0133]

[0134] 2) Model training

[0135] The mean square error between the predicted value and the true value of the heat load time series is used as the loss function of the four types of prediction models, and the mini-batch gradient descent algorithm and Adam optimization algorithm are selected for model training to minimize the loss function of the model.

[0136] For the recursive and direct prediction methods, the output of the prediction model is the heat load value at a certain moment in the prediction range, and the calculation formula of the loss function is shown in the following formula. The output of the LSTM-seq2seq prediction model and the TALSTM-seq2seq prediction model is the total heat load value in the prediction range, and the calculation formula of the loss function is shown in the following formula:

[0137]

[0138] Where θ represents all learnable parameters in the prediction model; K represents the batch size, that is, the number of training samples used to calculate gradients and update model parameters in one iterative training; It represents the true value of the heat load time series within the forecast range; Represents the predicted value of the heat load time series within the forecast range.

[0139] (3) Evaluation of prediction results

[0140] The mean absolute error (MAE) and root mean square error (CV-RMSE) are still used as evaluation indicators for prediction results. All prediction models are trained 10 times, and then the average of the prediction evaluation indicators of the model on the test set is calculated. The following is the average prediction result of each prediction model on the test set. In terms of prediction accuracy, the multi-input multi-output prediction method is the best, followed by the direct prediction method, and then the recursive prediction method. Moreover, among the multi-input multi-output prediction methods, the performance of the TALSTM-seq2seq prediction model is better than that of the LSTM-seq2seq prediction model.

[0141] Table 2 Average prediction results of each prediction model on the test set (the best prediction result is shown in bold)

[0142]

[0143] When predicting the short-term heat load profile, the recursive prediction method has the simplest modeling process and only needs to establish a single-step prediction model. However, the prediction of the heat load at the subsequent moment requires the prediction results of the previous moment. Considering the large range of heat load prediction in the case analysis, the recursive prediction method is prone to error accumulation. Moreover, the loss function of the recursive prediction model based on the LSTM network is the mean square error between the predicted value and the true value of the heat load at the next moment, rather than the mean square error between the predicted value and the true value of all heat loads within the prediction range, that is, the model training goal and the goal of multi-step heat load prediction are inconsistent. Due to the above two problems of error accumulation and the deviation of the model training goal from the prediction goal, the recursive prediction method has the lowest prediction accuracy.

[0144] In order to predict the hourly heat load profile for the next day, the direct prediction method establishes 24 independent single-step prediction models, each of which is only responsible for the heat load prediction task at a certain moment within the prediction range. Although the loss function of each prediction model is only the mean square error between the predicted value and the true value of the heat load at a certain moment, after the integration of the 24 prediction models, the modeling goal of the direct prediction method is consistent with the goal of multi-step prediction of heat load. Therefore, the performance of the direct prediction model based on the LSTM network is better than the recursive prediction model based on the LSTM network.

[0145] The multi-input multi-output prediction method combines the advantages of recursive and direct prediction methods. By adopting the encoder-decoder architecture as the basic structure of the prediction model, the output of the model is all the heat loads within the prediction range. It also ensures that the mean square error between the predicted value and the true value of all heat loads within the prediction range can be used as the loss function of the model, so end-to-end model training can be achieved. The prediction results in Table 3.2 also confirm that the LSTM-seq2seq prediction model and TALSTM-seq2seq prediction model established using the multi-input multi-output prediction method have the best performance.

[0146] In addition, although using LSTM networks for encoding and decoding separately is an effective method to achieve multi-input multi-output prediction, when the input time series is too long, it is difficult to save all the information of the input sequence only through the LSTM network. The TALSTM-seq2seq prediction model adds a temporal attention mechanism layer in its decoder part, which can more effectively model the dynamic temporal patterns in the time series by adaptively selecting the hidden state of the encoder at each moment. Due to the use of the temporal attention mechanism, the TALSTM-seq2seq prediction model can more accurately predict the short-term heat load profile compared to the LSTM-seq2seq prediction model.

[0147] Fig.12 Excerpts of the hourly heat load change curves output by each prediction model are given. Except for the recursive prediction model based on the LSTM network, the hourly heat load change curves generated by the other three types of prediction models are basically consistent with the actual heat load change curve. Only there is a significant deviation at the trough of the heat load curve. The prediction accuracy meets the requirements of engineering applications.

[0148] The results of the above analysis show that the short-term heat load profile of the heating system is predicted by using the heat load multi-step prediction model. By using the LSTM network to build an encoder-decoder architecture and introducing the temporal attention mechanism, the TALSTM-seq2seq prediction model can obtain the best prediction performance, indicating that the heat load multi-step prediction model has the best prediction performance.

[0149] Although the preferred embodiments of the present disclosure have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments, and all changes and modifications falling within the scope of the present disclosure.

[0150] Obviously, those skilled in the art can make various changes and modifications to the present disclosure without departing from the spirit and scope of the present disclosure. Thus, if these modifications and variations of the present disclosure fall within the scope of the claims of the present disclosure and their equivalents, the present disclosure is also intended to include these modifications and variations.

[0151] Finally, it should be noted that the above implementation modes are only used to illustrate the technical solutions of the present disclosure rather than to limit its protection scope. Although the present disclosure is described in detail with reference to the above implementation modes, ordinary technicians in the relevant field should understand that after reading the present disclosure, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation modes of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the disclosed claims to be approved.

Claims

1. A multi-step prediction method for heat load based on multi-input, the method is implemented based on the following system, the system includes a heating device and a heating system, characterized in that: The method comprises: S1. Obtaining multivariate time series data of the heating system; S2. establishing a data set based on the multivariate time series data, and normalizing the data set to obtain a normalized data set; the normalized data set includes a training set; S3. A method for constructing a multi-step thermal load prediction model; the multi-step thermal load prediction model comprises a sequence-to-sequence model, an encoder, a decoder and a temporal attention mechanism; the input data of the multi-step thermal load prediction model comprises the training set; S4. Obtaining heat load multi-step prediction data using the heat load multi-step prediction model.

2. The multi-step prediction method for heat load based on multi-input according to claim 1 is characterized in that: The steps of S1 include: S11. Constructing a directed connection graph based on the heating system; S12. Constructing a steady-state hydraulic simulation model based on the directed connection graph; the steady-state hydraulic simulation model simulates the operating state of the heating system; S13. Construct the multivariate time series data based on the operating status; the multivariate time series data includes target time series data and external cause time series data.

3. The multi-step prediction method for heat load based on multi-input according to claim 1 is characterized in that: The steps of S4 include: S41. Establishing a seq2seq sequence-to-sequence model; inputting the time series data into the seq2seq sequence-to-sequence model to obtain heat load data at all times within the prediction range; S42. Inputting the heat load data at all times within the prediction range into the encoder, using the encoder to obtain the average value of the hidden state at the current time, and obtaining the weighted hidden state of the encoder; S43. Using the decoder to obtain multi-step prediction data of heat load; S44. The decoder introduces a temporal attention mechanism to obtain an updated weighted hidden state of the encoder; the temporal attention mechanism includes an attention weight and a weighted sum; S45. Use the updated encoder weighted hidden state as the input of the decoder, and use the S43 method to obtain multi-step prediction data of thermal load.

4. The multi-step prediction method for heat load based on multi-input according to claim 4 is characterized in that: The weighted hidden state of the encoder in S42 is: In the formula, represents the weighted hidden state of the encoder, where D represents the number of neurons in the hidden layer of the encoder; t represents the time; τ represents the prediction range; h t =f e (h t-1 ,x t ), represents the hidden state at the current moment, where f e (·) is the LSTM neuron in the encoder; x is the external time series outdoor air temperature data; x t is the external time series at time t, h t-1 is the hidden state of the previous time step; Represents the hidden state of the encoder hidden layer at time t; W represents the length of the input training set and the external factor time series.

5. The multi-step prediction method for heat load based on multi-input according to claim 4 is characterized in that: The steps of S43 include: S431. Determine the hidden state of the decoder hidden layer, and when the hidden state of the decoder hidden layer is within the prediction range, predict the heat load; S432. Weight the encoder hidden state Input to the decoder and compared with the decoder hidden state d W+i-1 Splice to get the splicing vector S433. The concatenated vector is input to the fully connected layer, and after linear transformation, the heat load multi-step prediction data is output as: Where y is the target time series heat load data; Represents the predicted value of the target time series; Represents the linear transformation weight vector of the fully connected layer; represents the state of the fully connected layer, and D' represents the number of neurons in the fully connected layer; Represents the bias vector of the fully connected layer; represents the bias of the fully connected layer.

6. The multi-step prediction method for heat load based on multi-input according to claim 4 is characterized in that: The steps of S44 are as follows: S441. Get attention weight; set the model memory unit to The hidden state at the previous moment is d t'-1 ; The hidden state h of the encoder hidden layer at time t t With the memory cell c' t-1 Splice to get the splicing vector Obtain the correlation between the hidden state of the encoder hidden layer at time t and the concatenated vector: u t′t =V d T tanh(W d h t +U d [d t′-1 ;c t ″ -1 ]+b d )t=1,2,...,W+τ, In the formula, Represents the linear transformation weight vector of the temporal attention mechanism layer; is the input weight matrix, which represents the state of the temporal attention mechanism layer; is the state weight matrix, which represents the state of the temporal attention mechanism layer; Represents the bias vector of the temporal attention mechanism layer; S442. Perform weighted summation on the attention weight and the weighted hidden state of the encoder to obtain an updated weighted hidden state of the encoder: In the formula, β t't Represents the attention weight at the current moment.

7. A multi-step prediction device for heat load based on multi-input, the device is implemented based on the following system, the system includes a heating device and a heating system, characterized in that: The device comprises: Data acquisition module: used to acquire multivariate time series data of the heating system; Data preprocessing module: used to establish a data set based on the multivariate time series data, and also used to normalize the data set to obtain a normalized data set; the normalized data set includes a training set; Model building module: a method for building a heat load multi-step prediction model; the heat load multi-step prediction model includes a seq2seq sequence to sequence model, an encoder, a decoder and a temporal attention mechanism; the input data of the heat load multi-step prediction model includes the training set; Data prediction module: used to obtain heat load multi-step prediction data using the heat load multi-step prediction model.

8. A control device for a multi-step prediction method of heat load based on multi-input, the device comprising a processor and a memory, wherein: The memory is used to store computer programs, and is characterized in that when the microprocessor reads the computer program on the memory, it executes a multi-step prediction method for heat load based on multi-input as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program implements a multi-step prediction method for heat load based on multi-input as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of a multi-step prediction method for heat load based on multi-element inputs according to claims 1 to 6 are implemented.

Images