A Hierarchical Model for Summer Load Forecasting with Temperature-Sensitive Loads

Abstract

A hierarchical model for summer load forecasting with temperature-sensitive loads, which is applied to the day-ahead load forecasting problem of the summer power system. The steps are as follows: 1) Establish the first-layer neural network model to separate and predict the temperature-sensitive trend component in the system load. 2) Obtain the historical temperature-sensitive trend load and the historical residual load. 3) Establish and train the second-layer neural network. 4) Obtain the predicted load result. The present invention constructs a hierarchical neural network model. The first layer first trains a neural network considering the non-linear influence of temperature on the load to highlight the temperature influence; the temperature-sensitive trend component obtained from this layer is combined with the system load to obtain the residual load and sent to the second layer, considering other factors affecting the load other than temperature, to realize the prediction of the residual load; finally, the prediction results of the two layers are integrated to complete the load forecasting in summer. The idea of the present invention is simple and flexible, which can significantly improve the prediction accuracy of the model and solve the problem that the primary influence of temperature is not highlighted when using neural networks for summer load forecasting.

Description

Technical Field

[0001] The present invention belongs to the field of power system load forecasting, and particularly relates to a load forecasting model and its application to the load forecasting field of summer with temperature-sensitive loads. Background Art

[0002] With the continuous growth of power demand, the load of the power system in summer is getting larger and larger. During peak load periods, the accuracy of load forecasting is relatively poor. The poor accuracy has adverse consequences for the operation of the power system in many aspects. In the long term, the maximum load cannot be accurately predicted, making it difficult to reasonably arrange maintenance plans; in the short term, the load changes on the next day cannot be clearly analyzed, and it is impossible to appropriately arrange power generation plans. Therefore, the accuracy of load forecasting will directly affect the economy and security of power grid operation. Therefore, it is necessary to improve the accuracy of summer load forecasting.

[0003] Load forecasting needs to consider meteorological factors, date factors, and historical load factors. Meteorological factors include temperature, wind, etc.; date factors refer to Sunday factors; historical load factors refer to the influence of the load in the previous few days. From a mathematical perspective, the load forecasting problem is a multi-variable non-linear regression problem. The neural network model has strong advantages in dealing with multi-variable non-linear regression problems and has become the mainstream forecasting model in the field of load forecasting. However, the influence of these factors affecting the load on the load is not the same, and the neural network does not consider the differences between different factors in terms of structure when considering different factors. Summer temperature is the primary factor affecting the load, and its influence is much greater than other factors. It is necessary to highlight its influence. However, the existing neural network models do not consider highlighting the influence of temperature, resulting in a reduction in the accuracy of summer load forecasting.

[0004] The present invention constructs a hierarchical neural network model. In the first layer, a neural network considering the non-linear influence of temperature on the load is first trained to highlight the influence of temperature. The temperature-sensitive trend quantity obtained from this layer is combined with the system load to obtain the residual load and sent to the second layer. Considering other factors affecting the load other than temperature, the prediction of the residual load is realized. Finally, the prediction results of the two layers are integrated to complete the summer load forecasting. The forecasting accuracy of the present invention is relatively high. Summary of the Invention

[0005] Aiming at the problem that the primary influence of temperature is not highlighted when using a neural network for load forecasting, the present invention constructs a hierarchical neural network model. In the first layer, a neural network considering the non-linear influence of temperature on the load is first trained to highlight the influence of temperature. The temperature-sensitive trend quantity obtained from this layer is combined with the system load to obtain the residual load and sent to the second layer. Combining other factors affecting the load other than temperature, the prediction of the residual load is realized. Finally, the prediction results of the two layers are integrated to complete the summer load forecasting. The specific structure is as Figure 1as shown

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A hierarchical model for predicting summer loads with temperature-sensitive loads includes the following steps:

[0008] Step 1: Establish a first-layer neural network model to separate and predict the temperature-sensitive trend in the system load.

[0009] The first-layer neural network mentioned above needs to explore the relationship between temperature and load. Temperature has both an immediate impact on load, that is, the impact of real-time temperature on load, and a temperature cumulative effect, that is, the impact of temperature data in previous days. That is, when the temperature is continuously high in previous days, although the temperature remains unchanged, the load still shows a growth pattern that is fast first and then slow. In essence, the impact of temperature on load is a multiple non-linear regression problem. Therefore, the model requires real-time and temperature data in previous days as the input of the model. In addition, since the relationship between temperature and load at each moment is not exactly the same, hour information also needs to be introduced to distinguish the relationship between temperature and load at different moments. Based on the above analysis, according to the characteristics of the TCN neural network, the input and output of the first-layer neural network are obtained as shown in Equation (1).

[0010]

[0011] where, T i represents the temperature i-1 hours before the load to be fitted, h i represents the number of hours i-1 hours before the load to be fitted, LT represents the temperature trend load to be fitted, and is replaced by the system load L during training.

[0012] After the input and output of the model are determined, the first-layer neural network in the hierarchical model is established. The input layer is responsible for reading the input in Equation (1). The number of filters in the convolutional kernel of the TCN layer is set to 128, the time step parameter is set to 72, and the output of the Dense layer is set to 1, as Figure 2 shown

[0013] Step 2: Obtain the historical temperature-sensitive trend load and the historical residual load. By using the data in a period of time before the day to be predicted as the training set, the first-layer neural network established in Step 1 is trained to obtain the historical temperature-sensitive trend load in the historical system load. Then, subtracting the historical temperature-sensitive trend load from the historical system load can obtain the historical residual load.

[0014] Step 3: Establish and train a second-layer neural network to consider the effect of other factors on the load besides temperature.

[0015] In addition to temperature, there are many factors that affect the load, such as non-temperature influencing factors like wind power, historical load, and prediction day type. The first-layer neural network explores the relationship between temperature and load. After subtracting the temperature-influenced trend load output by the first-layer neural network from the total load, the prediction of the system load can be transformed into the prediction of the residual load, thus eliminating the influence of temperature on load prediction and studying the role of non-temperature influencing factors. That is, the second-layer neural network only needs to focus on the influence of non-temperature influencing factors on the residual load. Based on the LSTM neural network, the second-layer neural network is built to complete the prediction of the residual load.

[0016] The second-layer neural network needs to consider the relationship between other factors besides temperature and the load, including non-temperature meteorological factors such as wind power, historical load factors, and prediction day type factors. Since the load at different hours has different relationships with different factors, hours also need to be introduced as an input. Based on the above analysis, the input and output of the second-layer neural network are established according to the characteristics of the LSTM neural network as shown in Equation (2).

[0017]

[0018] Among them, R i represents the residual load 23 + i hours before the time point to be predicted, W i represents the wind speed i - 1 hours before the prediction time point, h i represents the hour number i - 1 hours before the time point to be predicted, represents the one-hot encoded week data i - 1 hours before the time point to be predicted, and R represents the residual load to be predicted.

[0019] After the input and output of the model are determined, the second-layer neural network in the hierarchical model is established. The input layer is responsible for reading the input in Equation (2). The number of neurons in the LSTM layer is set to 128, the time step parameter is set to 72, and the output of the Dense layer is set to 1, as Figure 3 shown.

[0020] After determining the specific structure of the second-layer neural network, the second-layer neural network is trained. The training set of the second-layer neural network is derived from the result of the first-layer neural network. First, the training set of the first-layer neural network needs to be prepared, and then the first-layer neural network is trained. That is, after obtaining the training result of the first-layer neural network, the input required by the second-layer neural network as shown in Equation (2) is added to obtain the training set of the second-layer neural network; then the second-layer neural network is trained. The training process is as Figure 4 shown. After the above training process, the trained second-layer neural network can be obtained.

[0021] Step 4: Obtain the predicted load result. Use the first-layer neural network to obtain the temperature trend load for the day to be predicted, and use the second-layer neural network to obtain the residual load for the day to be predicted. Add the two to obtain the system load for the day to be predicted. The specific prediction process is as Figure 5 shown.

[0022] The beneficial effects of the present invention are as follows:

[0023] The present invention solves the problem that the primary influence of temperature is not highlighted when using neural networks for summer load prediction, and its calculation speed is relatively fast. Traditional neural networks do not consider the differences of different factors in their structures, which causes the importance of temperature, the most important influencing factor for summer load, to be ignored. The present invention constructs a hierarchical neural network model. In the first layer, a neural network considering the non-linear influence of temperature on load is first trained to highlight the influence of temperature. The temperature-sensitive trend obtained from this layer is combined with the system load to obtain the residual load and sent to the second layer. Considering other factors that affect the load but are not temperature-related, the prediction of the residual load is realized. Finally, the prediction results of the two layers are integrated to complete the summer load prediction. The present invention considers the primary and secondary roles of different factors in summer short-term load prediction, highlights the influence of temperature, and improves the accuracy of summer short-term load prediction. Description of the Drawings

[0024] Figure 1 is the structure diagram of the hierarchical neural network;

[0025] Figure 2 is the structure diagram of the first-layer neural network;

[0026] Figure 3 is the structure diagram of the second-layer neural network;

[0027] Figure 4 is the training flow chart of the second-layer neural network;

[0028] Figure 5 is the prediction flow chart of the hierarchical model;

[0029] Figure 6 is the prediction result diagram of the hierarchical model. Detailed Embodiment

[0030] Aiming at the problem that the primary influence of temperature is not highlighted when using neural networks for load prediction, the present invention constructs a hierarchical neural network model. In the first layer, a neural network considering the non-linear influence of temperature on load is first trained to highlight the influence of temperature. The temperature-sensitive trend obtained from this layer is combined with the system load to obtain the residual load and sent to the second layer. Considering other factors that affect the load but are not temperature-related, the prediction of the residual load is realized. Finally, the prediction results of the two layers are integrated to complete the summer load prediction. The specific structure is as Figure 1As shown in the figure. Taking the data from June to July 2018 in a certain region of our country as the training set to predict the load from August 1st to 7th, 2018 as an example.

[0031] A hierarchical model for summer load forecasting with temperature-sensitive loads, comprising the following steps:

[0032] Step 1: Establish the first-layer neural network model to separate and predict the temperature-sensitive trend in the system load. The first-layer neural network needs to explore the relationship between temperature and load. Temperature has both an immediate impact on load, that is, the impact of real-time temperature on load, and a temperature cumulative effect, that is, the impact of temperature data from previous days. That is, when the temperature has been continuously high in the previous days, although the temperature remains unchanged, the load still shows a growth pattern that is fast first and then slow. In essence, the impact of temperature on load is a multiple non-linear regression problem. Therefore, the model requires real-time and temperature data from previous days as the input of the model. In addition, since the relationship between temperature and load at each moment is not exactly the same, hour information also needs to be introduced to distinguish the relationship between temperature and load at different moments. Based on the above analysis, the input and output of the first-layer neural network are obtained according to the characteristics of the TCN neural network as shown in Equation (1).

[0033]

[0034] output = LT

[0035] where, T i represents the temperature i - 1 hours before the load to be fitted, h i represents the number of hours i - 1 hours before the load to be fitted, LT represents the temperature trend load to be fitted, and is replaced by the system load L during training. Considering the temperature cumulative effect within 3 days, all temperatures within 3 days are used as the input of the model. Since hourly temperature data is used, i = 3 * 24 is taken.

[0036] Part of the input and output of the training set of the first-layer neural network is shown in Table 1.

[0037] Table 1 Part of the input and output of the training set of the first-layer neural network

[0038]

[0039]

[0040] After the input and output of the model are determined, the first-layer neural network in the hierarchical model is established, where the input layer is responsible for reading the input in Equation (1). The number of filters in the convolutional kernel of the TCN layer is set to 128, the time step parameter is set to 72, and the output of the Dense layer is set to 1, as Figure 2 shown.

[0041] Step 2: Obtain the historical temperature-sensitive trend load and the historical residual load. By using the data of a period before the day to be predicted as the training set to train the first-layer neural network established in Step 1, the historical temperature-sensitive trend load in the historical system load can be obtained. Then, subtracting the historical temperature-sensitive trend load from the historical system load can obtain the historical residual load. Part of the historical data is shown in Table 2.

[0042] Table 2 Part of the historical data

[0043]

[0044] Step 3: Establish and train the second-layer neural network to consider the effects of other factors besides temperature on the load.

[0045] Besides temperature, there are many factors affecting the load, such as non-temperature influencing factors like wind power, historical load, and the type of the day to be predicted. The first-layer neural network explores the variation relationship between temperature and load. After subtracting the temperature influence trend load output by the first-layer neural network from the total load, the prediction of the system load can be transformed into the prediction of the residual load, thereby eliminating the influence of temperature on the load prediction and studying the effects of non-temperature influencing factors. That is, the second-layer neural network only needs to focus on the influence of non-temperature influencing factors on the residual load. Build the second-layer neural network based on the LSTM neural network to complete the prediction of the residual load.

[0046] The second-layer neural network needs to consider the relationship between other factors besides temperature and the load, including non-temperature meteorological factors such as wind power, historical load factors, and the type of the day to be predicted factors. Since the relationship between the load at different hours and different factors is different, the hour also needs to be introduced as an input. Based on the above analysis, the input and output of the second-layer neural network are established according to the characteristics of the LSTM neural network as shown in Equation (2).

[0047]

[0048] output = R

[0049] where, R i represents the residual load i hours before the time point 23 + i to be predicted, W i represents the wind speed i - 1 hours before the prediction time point, h i represents the hour number i - 1 hours before the time point to be predicted, represents the one-hot encoded week data i - 1 hours before the time point to be predicted, R represents the residual load to be predicted, considering the residual load of each hour within 3 days, i = 3 * 24.

[0050] Part of the input and output of the second-layer neural network is shown in Table 3.

[0051] Table 3 Partial Input and Output of the Second - layer Neural Network Training Set

[0052]

[0053] After determining the input and output of the model, the second - layer neural network in the hierarchical model is established. The input layer is responsible for reading the input in formula (2). The number of neurons in the LSTM layer is set to 128, the time - step parameter is set to 72, and the output of the Dense layer is set to 1, as Figure 3 shown.

[0054] After determining the specific structure of the second - layer neural network, train the second - layer neural network. First, the training set of the second - layer neural network comes from the result of the first - layer neural network. It is necessary to prepare the training set of the first - layer neural network first, and then train the first - layer neural network. Then, after obtaining the training result of the first - layer neural network, add other non - temperature factors required by the second - layer neural network to obtain the training set of the second - layer neural network, as shown in Table 3. Finally, train the second - layer neural network. The training process is as Figure 4 shown. After the above training process, the trained second - layer neural network can be obtained.

[0055] Step 4: Obtain the predicted load result. Use the first - layer neural network to obtain the temperature - trend load of the day to be predicted, and use the second - layer neural network to obtain the residual load of the day to be predicted. Add the two to obtain the system load of the day to be predicted. The specific prediction process is as Figure 5 shown.

[0056] According to the above training process, the trained first - layer neural network and second - layer neural network are obtained, and the load from August 1, 2018 to August 7, 2018 is predicted. The specific prediction results are as Figure 6 shown. Among them, part of the prediction results are shown in Table 4.

[0057] Table 4 Part of the Prediction Results

[0058]

[0059] Conduct statistical analysis on the prediction results. The specific results are shown in Table 5.

[0060] Table 5 Statistical Analysis Table of Prediction Results

[0061]

[0062] The present invention solves the problem that the primary influence of temperature is not highlighted when using a neural network for load forecasting in summer, and its calculation speed is relatively fast. The traditional neural network does not consider the differences of different factors in its structure, resulting in the neglect of the importance of temperature, which is the most important factor in summer. The present invention constructs a hierarchical neural network model. In the first layer, a neural network considering the non-linear influence of temperature on load is trained first to highlight the influence of temperature. The temperature-sensitive trend quantity obtained from this layer is combined with the system load to obtain the residual load and sent to the second layer, where other factors affecting the load other than temperature are considered to realize the prediction of the residual load. Finally, the prediction results of the two layers are integrated to complete the load forecasting in summer.

[0063] As described above, the embodiments only represent the implementation manners of the present invention, but should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A hierarchical model for summer load forecasting with temperature-sensitive loads, characterized in that, It includes the following contents: Step 1: Establish a first-layer neural network model to separate and predict the temperature-sensitive trend in the system load; The first-layer neural network explores the relationship between temperature and load, and obtains the input and output of the first-layer neural network as shown in Equation (1) according to the characteristics of the TCN neural network; output = LT Among them, T i represents the temperature of the load to be fitted one hour before i, h i represents the number of hours one hour before the load to be fitted i, and LT represents the temperature trend load to be fitted, which is replaced by the system load L during training; After the input and output of the model are determined, the first-layer neural network in the hierarchical model is established; Step 2: Obtain the historical temperature-sensitive trend load and historical residual load; by using the data of a period of time before the day to be predicted as the training set, the first-layer neural network established in Step 1 can be trained to obtain the historical temperature-sensitive trend load in the historical system load, and then subtracting the historical temperature-sensitive trend load from the historical system load can obtain the historical residual load; Step 3: Establish and train a second-layer neural network to consider the effect of non-temperature factors on the load; The first-layer neural network explores the change relationship between temperature and load. After subtracting the temperature influence trend load output by the first-layer neural network from the total load, the prediction of the system load is transformed into the prediction of the residual load, so as to eliminate the influence of temperature on the load prediction and study the effect of non-temperature influence factors. That is, the second-layer neural network only needs to focus on the influence of non-temperature influence factors on the residual load; build the second-layer neural network based on the LSTM neural network to complete the prediction of the residual load; establish the input and output of the second-layer neural network as shown in Equation (2) according to the characteristics of the LSTM neural network; output = R Among them, R i represents the residual load 23 + i hours before the time point to be predicted, W i represents the wind speed i - 1 hours before the prediction time point, h i represents the number of hours i - 1 hours before the time point to be predicted, represents the one - hot encoded week data i - 1 hours before the time point to be predicted, and R represents the residual load to be predicted; After the input and output of the model are determined, the second-layer neural network in the hierarchical model is established; and the second-layer neural network is trained. The training set of the second-layer neural network is the training result of the first-layer neural network; Step 4: Obtain the predicted load result; use the first-layer neural network to obtain the temperature trend load of the day to be predicted, and use the second-layer neural network to obtain the residual load of the day to be predicted; add the two to obtain the system load of the day to be predicted.

Images