A comprehensive energy system multi-element load prediction method and system based on QR-TCN-LSTM

Abstract

The application discloses a kind of QR-TCN-LSTM-based comprehensive energy system multivariate load prediction method and system, and relates to load prediction technical field.The method includes obtaining a variety of load data with obvious time correlation, including power load data, heat load data and cold energy load data;The correlation coefficient between each load data and environmental influence factor is calculated respectively, and the key environmental influence factor is determined according to the correlation coefficient;According to the load data and key environmental influence factor, a data set is constructed, and the data set is divided into training set and test set;QR-TCN-LSTM model is constructed, and QR-TCN-LSTM model is trained using training set, to obtain trained QR-TCN-LSTM model;Load probability prediction is carried out using trained QR-TCN-LSTM model, to obtain the probability prediction result of each load data under different confidence intervals.Through the above method and system, the prediction accuracy and generalization ability of multivariate load are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of load prediction, more particularly to a comprehensive energy system multi-element load prediction method and system based on QR-TCN-LSTM. BACKGROUND

[0002] The integrated energy system (IES) can integrate electricity, heat, cold energy and renewable energy, etc., to realize efficient utilization and complementarity of resources. Multi-element load prediction, as a key technology for IES planning and operation, is of great significance to improving the reliability and economy of energy systems. However, due to the complex coupling relationship between multiple energy forms and the diversity and uncertainty of user demand, accurate load prediction faces severe challenges, so it is extremely crucial to study advanced multi-element load prediction methods for optimizing the scheduling and energy storage management of IES.

[0003] Traditional load prediction methods mainly include statistical methods, machine learning methods and deep learning methods. Statistical methods such as regression analysis and time series analysis perform well in dealing with linear relationships, but are not up to the task when faced with complex nonlinear relationships. Machine learning methods such as support vector machines and random forests have improved prediction accuracy, but still fall short when dealing with multi-energy coupling and uncertainty. Deep learning methods such as LSTM and CNN have made significant progress in load prediction, but they mainly focus on point prediction and fail to effectively quantify prediction uncertainty, making it difficult to meet the demand of IES for load probability distribution.

[0004] In addition, existing methods often ignore the mutual influence between different energy loads, resulting in inaccurate prediction results.

[0005] Therefore, how to develop a multi-element load probability prediction method and system that can effectively capture the coupling relationship between multiple energy forms and quantify prediction uncertainty has become a hot and difficult research topic. SUMMARY

[0006] Therefore, the present application provides a comprehensive energy system multi-element load prediction method and system based on QR-TCN-LSTM for predicting multi-element load.

[0007] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0008] The present application discloses a comprehensive energy system multi-element load prediction method based on QR-TCN-LSTM, comprising the following steps:

[0009] Obtain multiple load data with obvious time correlation, including power load data, heat load data and cold energy load data;

[0010] calculate the correlation coefficient between each load data and environmental impact factor respectively, and determine the key environmental impact factor according to the correlation coefficient;

[0011] According to the load data and the key environmental impact factor, a data set is constructed, and the data set is divided into a training set and a test set;

[0012] The QR-TCN-LSTM model is constructed, and the QR-TCN-LSTM model is trained using the training set to obtain the trained QR-TCN-LSTM model;

[0013] The trained QR-TCN-LSTM model is used for load probability prediction to obtain the probability prediction result of each load data under different confidence intervals.

[0014] Further, in the step of calculating the correlation coefficient between each load data and environmental impact factor, the Pearson correlation coefficient between each load data and environmental impact factor is calculated as the final correlation coefficient.

[0015] Further, in the step of constructing the QR-TCN-LSTM model, the QR-TCN-LSTM model includes a TCN network layer, an LSTM network layer and a quantile regression prediction layer.

[0016] The TCN layer is used to receive the load data and the key environmental impact factor in the training set, and output a feature sequence containing local features and time-dependent relationships.

[0017] The LSTM network layer is used to process the feature sequence output by the TCN layer to obtain an LSTM network layer prediction result.

[0018] The quantile regression prediction layer receives the LSTM network layer prediction result and obtains the final probability prediction result of each load data under different confidence intervals by minimizing the Pinball loss function.

[0019] Further, the Pinball loss function specifically includes the following loss function:

[0020]

[0021] where L q,t represents the loss value of the target quantile q at time t, represents the q quantile of the LSTM network layer prediction result at time t, y t represents the true value at time t.

[0022] Further, in the training step of the QR-TCN-LSTM model by using the training set, the parameter of the QR-TCN-LSTM model is further optimized by using the gradient descent method.

[0023] The application further discloses a QR-TCN-LSTM-based comprehensive energy system multi-element load prediction system, which comprises:

[0024] The data acquisition module is used for acquiring a plurality of load data with obvious time correlation, including power load data, heat load data and cold energy load data.

[0025] The correlation coefficient calculation module is used for calculating the correlation coefficient between each load data and an environmental influence factor, and determining a key environmental influence factor according to the correlation coefficient.

[0026] The data processing module is used for constructing a data set according to the load data and the key environmental influence factor, and dividing the data set into a training set and a test set.

[0027] The model training module is used for constructing a QR-TCN-LSTM model, training the QR-TCN-LSTM model by using the training set, and obtaining a trained QR-TCN-LSTM model.

[0028] The load prediction module is used for performing load probability prediction by using the trained QR-TCN-LSTM model, and obtaining a probability prediction result of each load data under different confidence intervals.

[0029] Preferably, in the correlation coefficient calculation module, the Pearson correlation coefficient between each load data and an environmental influence factor is calculated as the final correlation coefficient.

[0030] Preferably, in the model training module, the constructed QR-TCN-LSTM model comprises a TCN network layer, an LSTM network layer and a quantile regression prediction layer.

[0031] The TCN layer is used for receiving the load data and the key environmental influence factor in the training set, and outputting a feature sequence containing local features and time-dependent relationships.

[0032] The LSTM network layer is used for processing the feature sequence output by the TCN layer, and obtaining an LSTM network layer prediction result.

[0033] The quantile regression prediction layer receives the LSTM network layer prediction result, and obtains a final probability prediction result of each load data under different confidence intervals by minimizing a Pinball loss function.

[0034] Preferably, the Pinball loss function minimized by the quantile regression prediction layer specifically comprises the following loss function:

[0035]

[0036] wherein L q,t represents the loss value of the target quantile q at time t, represents the q-quantile of the prediction result of the LSTM network layer at time t, y t represents the true value at time t.

[0037] Preferably, the model training module further comprises: optimizing the parameters of the QR-TCN-LSTM model by using the gradient descent method.

[0038] According to the above technical solution, compared with the prior art, the application discloses a comprehensive energy system multi-element load prediction method and system based on QR-TCN-LSTM, which has the following beneficial effects:

[0039] The application analyzes the characteristics of multi-element load in depth, identifies key features and captures the mutual dependence between different energy loads, then captures the time dependence in the load data through the trained QR-TCN-LSTM, and generates probability prediction at different quantile levels through quantile regression, which significantly improves the prediction accuracy and generalization ability of multi-element load. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0041] Figure 1 The comprehensive energy system multi-element load prediction method based on QR-TCN-LSTM provided by the application is shown in the whole process schematic diagram.

[0042] Figure 2 The overall framework flowchart of the QR-TCN-LSTM model provided by the application is shown.

[0043] Figure 3 The dilated convolution structure schematic diagram of the TCN layer in the QR-TCN-LSTM model provided by the application is shown.

[0044] Figure 4 The residual structure schematic diagram of the QR-TCN-LSTM model provided by the application is shown.

[0045] Figure 5 A structure diagram of each LSTM unit in the QR-TCN-LSTM model provided by the present application is shown.

[0046] Figure 6 A calculation result diagram of the Yule correlation coefficient of the electric, cold, thermal load and environmental factors provided by the embodiment of the present application is shown.

[0047] Figure 7 A prediction result diagram of the electric load data obtained by the method provided by the embodiment of the present application is shown.

[0048] Figure 8 A prediction result diagram of the cold energy load data obtained by the method provided by the embodiment of the present application is shown.

[0049] Figure 9 A prediction result diagram of the thermal load data obtained by the method provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0051] The embodiment of the present application first discloses a QR-TCN-LSTM-based comprehensive energy system multi-element load prediction method, as shown in Figure 1 The method comprises the following steps:

[0052] Obtaining a plurality of load data with obvious time correlation, including electric load data, thermal load data and cold energy load data;

[0053] Calculating the correlation coefficient between each load data and the environmental influence factor respectively, and determining the key environmental influence factor according to the correlation coefficient;

[0054] Building a data set according to the load data and the key environmental influence factor, and dividing the data set into a training set and a test set;

[0055] Building a QR-TCN-LSTM model, and training the QR-TCN-LSTM model using the training set to obtain a trained QR-TCN-LSTM model;

[0056] Using the trained QR-TCN-LSTM model to perform load probability prediction to obtain the probability prediction result of each load data under different confidence intervals.

[0057] The steps of the present application are further described in detail as follows.

[0058] In the step of obtaining a plurality of load data with obvious time correlation, the multiple load data used in the step of the present application is from the integrated energy system. Through autocorrelation analysis on the data in the integrated energy system, the power load data, the heat load data and the cold energy load data with obvious time correlation with the current time are obtained as the load data in the present application.

[0059] Specifically, the autocorrelation coefficient formula used in the present application is as follows:

[0060]

[0061] In the formula, τ represents time delay, Cov(X t ,X t-τ ) represents the covariance between the time series data X t and X t-τ , Var(X t ) and Var(X t-τ ) represent the variance of the time series data X t and X t-τ respectively. The autocorrelation coefficient ranges from-1 to 1, 0 represents no correlation, positive value represents positive correlation, negative value represents negative correlation, and finally the power load data, the heat load data and the cold energy load data with autocorrelation coefficient in a certain threshold range are selected as the load data.

[0062] After obtaining the above load data, the correlation coefficient between each kind of load data and the environmental impact factor is calculated respectively, and the key environmental impact factor is determined according to the correlation coefficient. The Pearson correlation analysis is used in the embodiment of the present application to calculate the correlation coefficient between the environmental impact factors, and the calculation formula of the Pearson correlation coefficient is as follows:

[0063]

[0064] In the formula, x i represents the i-th sample in a certain kind of load data, y i represents the i-th sample in a certain kind of environmental data, and are the mean values of the load data x and the environmental data y respectively. The Pearson correlation coefficient is calculated by formula (2), and the environmental impact factors having a significant linear relationship with the load change are identified. Specifically, in the embodiments of the present application, the environmental impact factors having a Pearson correlation average absolute value greater than 0.3 with each kind of load data are selected as key environmental impact factors according to the results of the Pearson correlation analysis, and specifically include dew point, temperature, precipitation, air pressure and relative humidity.

[0065] According to the obtained load data and key environmental impact factors, a data set is constructed, and the data set is divided into a training set and a test set.

[0066] The present application realizes the prediction of load by constructing a QR-TCN-LSTM model, as shown in the following formula (3) : Figure 2 The QR-TCN-LSTM model includes a TCN network layer, an LSTM network layer and a quantile regression prediction layer.

[0067] Firstly, the TCN network layer receives the load data and the key environmental impact factors in the training set as the input layer of the QR-TCN-LSTM model, and converts the load data and the key environmental impact factors into one-dimensional time series data. The TCN network layer captures the local dependence relationship in the time series through dilated causal convolution. For one-dimensional time series input X = (x0, x1,..., x t ) and convolution kernel f, the extended convolution operation F(·) is defined as follows at sequence element t:

[0068]

[0069] Where d is the expansion rate, which determines the interval of the elements in the convolution kernel. By introducing the expansion rate d, the convolution kernel can expand the receptive field without increasing the computational cost. k is the size of the convolution kernel. The dilated convolution structure is as shown in the following formula (4) : Figure 3

[0070] With the increase of the number of TCN network layers, the problems of gradient disappearance and gradient explosion become more prominent. In order to solve these problems, TCN introduces residual connections. Residual connections allow the input signal to be directly transmitted to the output, so that the gradient can be more effectively propagated, thereby alleviating the problem of gradient disappearance in deep network. The residual structure is as shown in the following formula (5) : Figure 4 ​As shown, the residual structure is to connect the dilated causal convolution layer Dilated Causal Conv, the weight normalization layer WeightNorm, the ReLU activation function Dropout layer through the residual. Each residual block usually includes two branches, one branch is responsible for performing the conversion operation, which contains the dilated causal convolution layer, the weight normalization layer, the ReLU activation function and the Dropout layer. The right branch is a residual connection composed of a 1*1 convolution, and the two branches constitute a residual block in parallel. In order to ensure that the network can effectively train the deep architecture, the TCN layer uses the residual connection, assuming that the input of the hth residual block is X (h-1) , then the output X (h) is:

[0071] X (h) =σ(X (h-1) +f(X (h-1) )) (4)

[0072] Where σ(·) is the activation function.

[0073] Next, the LSTM network layer of the QR-TCN-LSTM model processes the output of the TCN layer to capture long-term dependencies. The key of the LSTM network layer lies in its unique cell structure, in the embodiment of the present application, the LSTM network layer is composed of multiple LSTM cell layers stacked, and the output of the previous LSTM cell layer is taken as the input of the next LSTM cell layer. This structure can learn more complex features and long-term dependencies in sequence data, and each LSTM cell contains three main gates: input gate, forget gate and output gate, and the basic unit of the LSTM cell network is as shown in Figure 5 The first gate in the LSTM cell is the forget gate f t , which determines how much information is retained from the memory cell state C t-1 at the previous moment. The state formula of the forget gate at time t is:

[0074] f t =σ(W f ·[h t-1 ,X t ]+b f ) (5)

[0075] Where σ(·) represents the sigmoid activation function; X t is the input at the current moment, mainly including load data and key environmental impact factors; h t-1 and b f represent the output vector at t-1 moment and the forget gate bias at t moment respectively; W f is the weight matrix of the forget gate.

[0076] Secondly, the input gate i t controls how much of the current information needs to be used as input to generate the current state C t . The formula for the input gate is:

[0077] i t = sigma(W i ·[h t-1 , X t ]+b i ) (6)

[0078] At the same time, the LSTM generates a new candidate memory cell state The formula for calculating it is:

[0079]

[0080] where W i and W c are the weight matrices of the input gate and the candidate memory cell state, b i and b c are the bias vectors, and tanh(·) is the activation function.

[0081] The update of the memory cell state C t is determined by the memory cell state C t-1 at the previous time and the current input information, and the formula is:

[0082]

[0083] The last stage of the LSTM unit is to calculate how much information can be used as output. Another control gate is used as the output gate o t :

[0084] o t = sigma(W o ·[h t-1 , X t ]+b o ) (9)

[0085] The hidden state h t is determined by the current output gate and the memory cell state, and the final hidden layer output h t of the LSTM is defined as:

[0086] h t = o t *tanh(C t ) (10)

[0087] In order to improve the computational efficiency and prediction accuracy of the model, the gradient descent method is used to optimize the model parameters during the model training process.

[0088] In the present application, the TCN layer and the LSTM network layer can be used as a TCN-LSTM model, and a quantile regression prediction module is connected after the output layer of the TCN-LSTM model. The core of the module is to map the hidden state vector h t to the predicted value at different quantile levels through linear transformation. For each quantile τ, define the trainable parameter weight vector W τ and the bias term b τ , and calculate the predicted value corresponding to the quantile τ through linear transformation, and the specific expression is as follows:

[0089] Q τ (Y|X)=W τ ·h t +b τ

[0090] where Q(Y|X) is the predicted value corresponding to the quantile τ; h t is the hidden state output of the TCN-LSTM model. Different quantile levels such as τ=0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95 are set to enable the model to output multivariate load prediction values Q τ1 (Y|X), Q τ2 (Y|X),..., and Q τk (Y|X) at different probability levels. The quantile regression layer parameters are trained by minimizing the Pinball loss function to measure the deviation of the model prediction value from the true value at different quantile levels, ensuring accurate prediction in each probability interval. The loss function is:

[0091]

[0092] where q represents the target quantile, L q,t q(t) represents the estimated q-th quantile at time t, and L q,t represents the loss value of the q-th quantile at time t. The smaller the loss value, the better the probability prediction effect.

[0093] After that, the quantile regression prediction layer is deeply integrated with the TCN-LSTM model. For any prediction model, it needs to be trained according to the appropriate loss function and optimization strategy. In the training stage, the parameters of the prediction model are obtained by minimizing the loss function:

[0094]

[0095] In the formula, f q (X t ,W q ) is the predicted value, and y t is the true value.

[0096] By dynamically adjusting the learning rate during training, the model is prevented from falling into a local optimal solution and the convergence speed is accelerated. By using the cross-validation method, the training data set is divided into multiple subsets, one subset is used as the validation set and the rest is used as the training set, the model is trained and evaluated for multiple rounds, the root mean square error (RMSE), average coverage error (ACE) and other evaluation indexes on the validation set are used to adjust the hyperparameters such as the number of neural network layers and nodes, and the optimal model performance is obtained. Finally, the preprocessed real-time multi-element load data is input into the trained QR-TCN-LSTM model, and the multi-element load prediction values at different quantile levels are output by the quantile regression layer. These prediction results in the form of probability distribution can help energy system decision makers to make scientific and reasonable energy production, distribution and scheduling plans, for example, energy suppliers can arrange power generation plans according to the power load prediction values at different probability levels, to avoid the problem of energy shortage or surplus caused by load fluctuation.

[0097] In order to realize the above method, the application further discloses a QR-TCN-LSTM-based comprehensive energy system multi-element load prediction system, which comprises:

[0098] A data acquisition module is configured to acquire a plurality of load data with obvious time correlation, including power load data, heat load data and cold energy load data.

[0099] A correlation coefficient calculation module is configured to calculate the correlation coefficient between each load data and environmental influence factors, and determine the key environmental influence factors according to the correlation coefficient.

[0100] A data processing module is configured to construct a data set according to the load data and the key environmental influence factors, and divide the data set into a training set and a test set.

[0101] A model training module is configured to construct a QR-TCN-LSTM model, and train the QR-TCN-LSTM model using the training set to obtain a trained QR-TCN-LSTM model.

[0102] A load prediction module is configured to perform load probability prediction using the trained QR-TCN-LSTM model to obtain the probability prediction results of each load data at different confidence intervals.

[0103] The specific implementation process of each module can refer to the description in the method, which will not be repeated here.

[0104] The effectiveness of the application is illustrated by specific examples as follows:

[0105] The load data source of the embodiment of the application is the historical data of the electric, cold and heat load of a certain region, and the weather data source is the weather data of the location of the region in the climate data center website, including temperature, atmospheric pressure, wind speed and wind direction. The sampling time of all data is from 0 o'clock on May 2, 2018 to 23 o'clock on December 31, 2019, and the sampling interval of all data is 1 hour. The application uses the above data to jointly predict the electric, cold and heat load of the region in the future one week, and the specific results and analysis are as follows:

[0106] 1. Load characteristic analysis

[0107] The application calculates the correlation coefficients between the collected cold, heat and electric loads and the environmental factors such as dew point, temperature, relative humidity, wind speed, precipitation and air pressure and the load. The Pearson correlation coefficient between each environmental factor and the load data is calculated as shown in Figure 6 .

[0108] It can be seen from Figure 6 that there is a strong correlation between the environmental factors and the different load data, which also shows that the coupling relationship between the multiple loads should be fully considered when performing load prediction in IES, and a multiple load joint prediction model should be used. In order to effectively train the model, the environmental factors that have a significant impact on the prediction performance are selected, when the absolute value of the correlation coefficient is greater than 0.3, it is considered that there is a strong correlation between the environmental factors and the load data, that is, the dew point, temperature, precipitation, air pressure and relative humidity are selected as the key environmental factors, and the electric, heat and cold energy three load data and the key environmental factors are jointly used as the input features of the QR-TCN-LSTM model.

[0109] 2. Comparison method setting

[0110] In order to verify the effectiveness and reliability of the model proposed in the application, the experimental results of the model are compared with those of other five probability prediction models, including: LSTM, recurrent neural network (RNN), GRU and time convolution network (TCN) four typical time series prediction models; CNN-LSTM model, which can effectively extract the coupling relationship and time sequence characteristics between multiple time series.

[0111] The hyperparameter setting is crucial to the realization of the good prediction performance of the model. The hyperparameters are determined by the experience method and the grid search method. First, according to the actual operation experience and the hyperparameters provided in the related literature, the initial values of the hyperparameters of the prediction model are determined. Then, the grid search method is used to find the optimal hyperparameters in the search range. Considering the limitation of machine memory and training speed, the batch size is set to 32 in the test, the model optimizer adopts the Adam method, and the learning rate is 0.0001. In order to ensure the credibility of the results, 5 repeated tests are performed on each model, and the average value of the evaluation index is taken as the final result.

[0112] To evaluate the performance of the constructed QR-TCN-LSTM multivariate load probability forecasting model, this invention selects Average Coverage Error (ACE), Average Width (AW), and Root Mean Squared Error (RMSE) as evaluation metrics. These metrics comprehensively reflect the model's forecasting performance in different aspects. ACE and AW are mainly used to measure the accuracy and width of the probability distribution, while RMSE is used to evaluate the overall accuracy and bias of the predicted values. The specific expressions for the evaluation metrics are as follows:

[0113]

[0114] in, This refers to the actual value y i Does it fall within the prediction range? Indicator functions within. and These represent the upper and lower limits of the prediction interval, respectively. It is the predicted value, y i These are the actual values. q is the confidence interval, and N is the sample size.

[0115] 3. Experimental Results and Analysis

[0116] To verify the effectiveness and reliability of the proposed model, this invention was compared with five other probabilistic prediction models, including: four typical time series prediction models: LSTM, Recurrent Neural Network (RNN), GRU, and Temporal Convolutional Network (TCN); and the CNN-LSTM model, which can effectively extract the coupling relationship and temporal features between multivariate time series.

[0117] To ensure the reliability of the results, this invention conducted five repeated experiments on each model, and took the average value of the evaluation index as the final result. The experiments were conducted at 50%, 70%, and 90% confidence intervals, and the evaluation results are shown in Table 1.

[0118] Table 1 Comparison of Evaluation Results of Different Models

[0119]

[0120] As can be seen from Table 1, under the 50% confidence interval, the ACE values of the method of the present application on the electrical load, cold load and thermal load are 0.0119, 0.0952 and 0.0416 respectively, which are significantly lower than those of other methods, indicating that the prediction interval can more accurately cover the true value, and indicating that the method of the present application has extremely high coverage accuracy at a lower confidence interval. In addition, the method of the present application also performs well in AW, and the AW values of the electrical load, cold load and thermal load are 1020.7, 1003.28 and 226.14 respectively, all of which are the smallest, showing that the prediction interval is the narrowest and the accuracy is the highest. Finally, the method of the present application also performs well in RMSE, and the RMSE values of the electrical load, cold load and thermal load are 582.75, 551.61 and 186.34 respectively, all of which are the smallest, indicating that the prediction error is the smallest and the performance is the best.

[0121] Under the 70% confidence interval, the advantage of the method of the present application is still significant. The ACE values of the method of the present application on the electrical load, cold load and thermal load are 0.1464, 0.1761 and 0.1781 respectively, which continue to maintain a leading position. At the same time, the AW values of the electrical load, cold load and thermal load are 1113.14, 1823.38 and 335.54 respectively, indicating that the prediction interval is relatively narrow and the accuracy is relatively high. In terms of RMSE, the RMSE values of the method of the present application on the electrical load, cold load and thermal load are 592.34, 530.75 and 155.88 respectively, all of which are the smallest, showing excellent prediction accuracy.

[0122] Under the 90% confidence interval, the ACE values of the method of the present application are 0.0761, 0.0821 and 0.0821 respectively, which are significantly better than those of other methods. The method of the present application still performs well in AW, and the AW values of the electrical load, cold load and thermal load are 4428.32, 3553.44 and 449.65 respectively, all of which are the smallest, indicating that the prediction interval is the narrowest and the accuracy is the highest. In terms of RMSE, the RMSE values of the method of the present application on the electrical load, cold load and thermal load are 553.75, 511.62 and 131.84 respectively, which have higher prediction performance.

[0123] Figure 7 The prediction results of the QR-TCN-LSTM model for electrical load under different confidence intervals are shown. As shown in Figure 7 , with the increase of the confidence interval, the prediction uncertainty also increases. Although there are significant peak fluctuations in part of the time period, the bias is mostly within the 90% confidence interval, showing the robustness and reliability of the model. The upper limit allowance is larger and the lower limit allowance is smaller, reflecting the conservatism of the model in dealing with peak load. Figure 8 The prediction results of the QR-TCN-LSTM model for cold energy load under different confidence intervals are shown. As shown in Figure 8As shown, the display shows the cold load change rule obviously, the model in the 50% confidence interval better captures the actual load change. Although in some sharp peak and trough turning, the predicted value deviates from the actual value, most of the cold load is still in the 90% confidence interval, which shows the prediction reliability of the model. The margin of the upper and lower limits of the confidence interval reflects the adaptability of the model in dealing with the cold load fluctuation. Figure 9 The prediction results of the QR-TCN-LSTM model for the thermal load under different confidence intervals are shown as follows: Figure 9 As shown, the heat load change trend is mild, but there are also fluctuations. With the increase of the confidence interval, the prediction interval becomes wider, covering more actual values, reflecting the uncertainty of the prediction results.

[0124] In summary, through the analysis of the test results under different confidence intervals, it can be seen that the method proposed in the present application shows the optimal prediction performance in the multi-element load probability prediction of the comprehensive energy system. Whether in the average coverage error (ACE), the average width (AW) or the root mean square error (RMSE), the method of the present application is superior to the other five comparison methods, further verifying the effectiveness and superiority of the method. This result fully proves the potential and practical value of the method in the multi-element load probability prediction.

[0125] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0126] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A QR-TCN-LSTM based comprehensive energy system multi-element load forecasting method, characterized in that, Includes the following steps: Acquire various load data with significant time correlation, including electricity load data, heat load data, and cooling load data; The correlation coefficient between each load data and environmental impact factors is calculated, and the key environmental impact factors are determined based on the correlation coefficients. The key environmental impact factors include dew point, temperature, precipitation, air pressure, and relative humidity. A dataset is constructed based on the load data and key environmental influencing factors, and the dataset is divided into a training set and a test set. Construct a QR-TCN-LSTM model and train it using a training set to obtain a trained QR-TCN-LSTM model; in the step of constructing the QR-TCN-LSTM model, the QR-TCN-LSTM model includes a TCN network layer, an LSTM network layer, and a quantile regression prediction layer; The TCN network layer is used to receive load data and key environmental influencing factors in the training set, and outputs a feature sequence containing local features and time dependencies. The LSTM network layer is used to process the feature sequence output by the TCN network layer to obtain the prediction result of the LSTM network layer. The quantile regression prediction layer receives the prediction results from the LSTM network layer and obtains the final probability prediction results for each load data under different confidence intervals by minimizing the Pinball loss function. The trained QR-TCN-LSTM model is used to predict the probability of load data, and the probability prediction results of each type of load data under different confidence intervals are obtained.

2. The QR-TCN-LSTM-based integrated energy system multi-element load forecasting method according to claim 1, characterized in that, The step of calculating the correlation coefficient between each load data and environmental impact factors specifically includes calculating the Pearson correlation coefficient between each load data and environmental impact factors, which is used as the final correlation coefficient.

3. The multi-element load forecasting method for integrated energy systems based on QR-TCN-LSTM according to claim 1, characterized in that, The Pinball loss function specifically includes the following loss functions: Among them, L q,t This represents the loss value of the target quantile q at time t. y represents the q-th quantile of the LSTM network layer prediction results at time t. t This represents the true value at time t.

4. The multi-element load forecasting method for integrated energy systems based on QR-TCN-LSTM according to claim 1, characterized in that, The training process for the QR-TCN-LSTM model using the training set also includes optimizing the parameters of the QR-TCN-LSTM model using gradient descent.

5. A multi-element load forecasting system for integrated energy systems based on QR-TCN-LSTM, characterized in that, include: The data acquisition module is used to collect various load data with obvious time correlation, including power load data, heat load data and cooling load data; The correlation coefficient calculation module is used to calculate the correlation coefficient between each type of load data and environmental impact factors, and to determine the key environmental impact factors based on the correlation coefficients; the key environmental impact factors include dew point, temperature, precipitation, air pressure, and relative humidity; The data processing module is used to construct a dataset based on the load data and key environmental influencing factors, and to divide the dataset into a training set and a test set. The model training module is used to construct the QR-TCN-LSTM model and train the QR-TCN-LSTM model using the training set to obtain the trained QR-TCN-LSTM model. The constructed QR-TCN-LSTM model includes a TCN network layer, an LSTM network layer, and a quantile regression prediction layer. The TCN network layer is used to receive load data and key environmental influencing factors in the training set, and outputs a feature sequence containing local features and time dependencies. The LSTM network layer is used to process the feature sequence output by the TCN network layer to obtain the prediction result of the LSTM network layer. The quantile regression prediction layer receives the prediction results from the LSTM network layer and obtains the final probability prediction results for each load data under different confidence intervals by minimizing the Pinball loss function. The load forecasting module is used to perform load probability forecasting using a trained QR-TCN-LSTM model, and obtain the probability forecast results for each type of load data under different confidence intervals.

6. The integrated energy system multi-element load forecasting system based on QR-TCN-LSTM according to claim 5, characterized in that, In the correlation coefficient calculation module, the Pearson correlation coefficient between each load data and environmental influencing factors is calculated as the final correlation coefficient.

7. A multi-element load forecasting system for integrated energy systems based on QR-TCN-LSTM according to claim 5, characterized in that, The pinball loss function minimized by the quantile regression prediction layer specifically includes the following loss functions: Among them, L q,t This represents the loss value of the target quantile q at time t. y represents the q-th quantile of the LSTM network layer prediction results at time t. t This represents the true value at time t.

8. A multi-element load forecasting system for integrated energy systems based on QR-TCN-LSTM according to claim 5, characterized in that, The model training module also includes optimizing the parameters of the QR-TCN-LSTM model using gradient descent.

Images