A distributed energy station heating load forecasting method based on LSTM

Abstract

The present invention discloses a distributed energy station heating load prediction method based on LSTM, which relates to the field of distributed energy station heating technology. The present invention proposes a distributed energy station heating load prediction method based on long short-term memory network. By learning the historical operating data, weather data, building data, and time data of each distributed energy station after exception processing by the isolation forest algorithm, the unique memory mechanism and gating structure of LSTM are applied to analyze the variation pattern of the heating load of the energy station, and the most reasonable outlet water temperature or return water temperature in the next hour is predicted based on real-time data. The temperature is promptly sent to the energy station to adjust the heating temperature of the air source heat pump, realizing precise control of heating, effectively maintaining room temperature balance, improving the COP value of the energy station, and reducing work orders and energy waste.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed energy station heating, and specifically to a distributed energy station heating load prediction method based on LSTM. Background Art

[0002] Distributed energy station heating involves placing heat source equipment near users and delivering heat directly to their devices via residential heating networks or internal building pipes. To maintain a consistent heating experience and minimize energy waste, distributed energy stations must promptly adjust heating loads based on factors such as weather, equipment conditions, and changes in heating area. Currently, load adjustment at distributed energy stations relies primarily on manual experience or configuration settings based on manual experience. The accuracy and efficiency of this adjustment process largely depends on the operator's expertise, accumulated experience, and keen insight into actual conditions.

[0003] Therefore, distributed energy station heating has certain limitations in load regulation: the scattered locations and widely varying conditions make unified regulation difficult and challenging. Relying on manual experience for decision-making makes it difficult to allocate sufficient manpower for precise regulation of each station, resulting in high labor costs. Adjustments are infrequent and untimely, failing to respond to subtle changes in actual load, leading to frequent under- or over-heating. Adjustments also rely heavily on manual expertise. Summary of the Invention

[0004] The purpose of the present invention is to provide a distributed energy station heating load forecasting method based on LSTM to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the heating load of a distributed energy station based on LSTM, the method comprising the following steps:

[0006] S1. Obtain historical data of a distributed energy station, perform data missing processing based on the historical data, and record the number of missing data processed;

[0007] S2. Build an isolation forest model based on historical data to identify abnormal data, perform data missing processing on the abnormal data, and record it in the missing processing quantity;

[0008] S3. For the processed data, a distributed energy station load forecasting model is constructed based on the long short-term memory network;

[0009] S4. Build a system evaluation model based on the real-time operation data of the distributed energy station load to realize the update and reuse of the distributed energy station load prediction model.

[0010] According to the above technical solution, the historical data includes: site operation data, weather data, building data and time data;

[0011] The site operation data includes supply water temperature, return water temperature, instantaneous heat, instantaneous flow, accumulated heat, accumulated electricity, indoor temperature and target room temperature; the weather data includes temperature, humidity, rainfall, wind speed, wind direction, weather type and solar radiation; the building data includes floor, heating area, building insulation level, building type and heating type; the time data includes the hour of the day, the week and whether it is a holiday.

[0012] According to the above technical solution, the data missing processing includes:

[0013] When historical data is missing, choose to abandon or linear interpolate the processing method, including:

[0014] If there are complete data records before and after the missing data point, the missing data is supplemented by linear interpolation. Linear interpolation estimates the value of the missing point by drawing a straight line based on the known data before and after the missing data point.

[0015] The drawing method includes:

[0016] Data exists before the missing point , data exists after the missing point , the time difference between the two existing data is , the time difference between the missing point and the data before the missing point is , then the missing point data is:

[0017]

[0018] If the data before and after the missing data point are incomplete, or if multiple data points are missing consecutively, choose to discard the missing data point.

[0019] According to the above technical solution, the abnormal data determination includes:

[0020] Grouping data into similar groups. The grouping process specifically includes:

[0021] Weather data: Weather codes and outdoor temperature are used as grouping criteria. Weather codes reflect changes in rainfall, humidity, and solar radiation. Outdoor temperature is divided into groups of 5 degrees Celsius, with the endpoints including the end point and excluding the end point.

[0022] Building data: take the number of floors as the grouping basis;

[0023] Time data: divided by the time range of the day, whether it is a weekend, and whether it is a holiday. The time range of the day is divided into daytime working hours and nighttime off-duty time. The "daytime working hours" and "nighttime off-duty time" are set according to the specific conditions of different sites. The default is 7:00 am to 7:00 pm and 7:00 pm to 7:00 am, where the endpoints are included before but not after.

[0024] According to the above technical solution, it also includes:

[0025] S5-1. For any group, randomly select N groups of heating operation data as subsamples and put them into the root node of an isolated tree;

[0026] S5-2. At each root node, randomly select a feature from the heating operation data and randomly select a value between the maximum and minimum values ​​of the feature as the split value. The data is divided into the left subtree or the right subtree according to this split value.

[0027] S5-3, randomly select one of the other data items of the heating operation data as a feature, and repeat step S5-2 until the data cannot be divided any further, resulting in an isolated tree;

[0028] S5-4. Repeat steps S5-1 to S5-3 to obtain M isolated trees to form an isolation forest, where M is the number of isolated trees obtained by applying the isolation forest algorithm to the data in one group;

[0029] S5-5. In the isolation forest, a threshold for the number of split points is set and any data is judged. During the isolation process, if the number of split points exceeds the threshold, it is judged as normal data; if the number of split points does not exceed the threshold, it is judged as abnormal data.

[0030] S5-6. Feedback the detected abnormal data for data missing processing.

[0031] According to the above technical solution, the distributed energy station load forecasting model includes:

[0032] Three gating units, namely the input gate, forget gate, and output gate, are introduced to control the dimensions of the heating data retained at the previous moment. LSTM input data includes: supply water temperature, return water temperature, instantaneous heat, instantaneous flow, indoor temperature, target room temperature, temperature, humidity, rainfall, wind speed, wind direction, weather type, solar radiation, floor, hour of the day, week number, and whether it is a holiday.

[0033] The forget gate determines the data information that LSTM discards from the heating data at the previous moment, specifically including:

[0034] Set the current time to t, and the forget gate reads the output of the previous layer and current input real-time data , calculate the retention factor of the information at the previous moment :

[0035]

[0036] Where: is the forget gate weight matrix, is the forget gate bias, For the sigmoid function, Mapped to the interval [0,1]; where, Generated by the input at the previous moment and the retained information of the previous moment;

[0037] The cell state obtained from the information of the previous moment and Multiply, A value of 1 indicates that All information in the , A value of 0 means completely discarding Information in;

[0038] The input gate is used to filter information and determine the new heating data input at the current moment to be used to update the final output data information, including a sigmoid layer and a tanh layer;

[0039] The sigmoid layer is based on the output of the previous layer. and current input real-time data Calculate the update factor of heating information at the new time :

[0040]

[0041] in: is the input gate sigmoid layer weight matrix, Bias for the input gate sigmoid layer;

[0042] The tanh layer calculates candidate cell state values , is added to the cell state at the current moment middle:

[0043]

[0044] in: is the input gate tanh layer weight matrix, is the input gate tanh layer bias, is the tanh activation function;

[0045] The current cell state It is equal to the sum of the product of the output value of the forget gate and the product of the two output values ​​of the input gate, that is:

[0046]

[0047] The output gate is used to calculate the output of the current layer ;

[0048] The output gate first passes through a sigmoid layer to determine the cell state at the current moment Output data information:

[0049]

[0050] Where: is the output gate weight matrix, is the output gate bias;

[0051] The current cell state By processing with tanh, a value in the interval [-1,1] is obtained and it is combined with Multiply and finally determine the output :

[0052]

[0053] If the current layer is the last layer, then the It is the final prediction value of the distributed energy station heating load prediction model.

[0054] According to the above technical solution, it also includes:

[0055] Set the prediction time span p and reference data time span k for the LSTM algorithm;

[0056] Assume that the current time is d, and use k data from d-k+1 to d to predict the energy station load value at the future p-th time;

[0057] When LSTM is required to provide the current load recommendation value: During the model training phase, set p = 0, and the training set consists of k data items. The target value of each data group is the load value corresponding to the kth data item. The weather data is the weather data at the corresponding moment of each data item. During the model inference phase, select k-1 historical data items and the current real-time data, combine them with the weather data at the corresponding moment, and apply the model to obtain the current load recommendation value.

[0058] When LSTM is needed to predict the load recommendation value at a future time, set the future time R. In the model training phase, set p = Rd. The training set consists of k data items. The target value of each group of data is the load value corresponding to the k+pth data item. The weather data is the weather data at the time + p corresponding to each data item. In the model inference phase, select k-1 historical data items and the current real-time data, combine them with the weather data at the time + p corresponding to each data item, and apply the model to predict the load recommendation value at the future time R.

[0059] According to the above technical solution, the system evaluation model includes:

[0060] Obtain the generated distributed energy station heating load forecasting model and the amount of missing data in the model construction process;

[0061] Define a model prediction tolerance threshold and perform an operational analysis on the generated distributed energy station heating load forecasting model. The operational analysis includes:

[0062] Starting from the current moment, obtain the absolute value of the difference between the actual result and the predicted result each time, and continue to accumulate it until the total absolute value of the difference exceeds the model prediction tolerance threshold for the first time. Record the total number of predicted data experienced from the current moment;

[0063] With the number of missing data as the horizontal axis and the total number of experienced forecast data as the vertical axis, a linear fit is performed on the coordinate points formed by the heating load forecasting model of all distributed energy stations to form a functional relationship between the number of missing data and the total number of experienced forecast data;

[0064] Obtain the number of missing data in the current distributed energy station heating load forecasting model, feedback the total number of forecast data experienced based on the functional relationship, mark the position, and when the model reaches the marked position, feedback is sent to the management port to remind the model to update.

[0065] According to the above technical solution, it also includes:

[0066] When a new site is put into use, its building attribute data is compared with that of the site in the existing heating load prediction model, and the sites that belong to the same group with its building attributes are calculated and reused. After the new site has been put into use for some time, the model is updated according to the historical operating data of the new site to ensure that the average room temperature of the site reaches a stable state.

[0067] Compared with the existing technology, the beneficial effects of the present invention are: the present invention can customize and generate its own unique load forecasting model based on the historical operating data and building data of different sites, greatly improving the accuracy of regulation. When the weather changes, the system can respond quickly and adjust the load to ensure that the average room temperature remains stable, reducing the number of work orders while effectively preventing energy waste. During special periods, such as nights, weekends or holidays, the system can implement specific temperature control strategies to further reduce energy consumption. At the same time, it can also intelligently manage the start-up time of the heat pump to avoid energy waste caused by premature start-up and heating. The prediction time span can be flexibly adjusted, and the load value can be predicted a few hours in the future, making the regulation time more flexible. It reduces dependence on manual operations, reduces labor costs, and improves overall operation and maintenance efficiency. When a new site is put into operation, there is no need to spend a lot of time on manual adjustment and optimization. The model can reuse data from other similar sites and combine it with the current site operation data to continuously update the model so that the site room temperature average reaches a stable state faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of a method for predicting the heating load of a distributed energy station based on LSTM according to the present invention;

[0069] Figure 2 This is a schematic diagram of the LSTM structure of a distributed energy station heating load prediction method based on LSTM in the present invention;

[0070] Figure 3 This is a schematic diagram of a forget gate in an embodiment of a method for predicting heating load of a distributed energy station based on LSTM according to the present invention;

[0071] Figure 4 Schematic diagram of an input gate of an embodiment of a method for predicting heating load of a distributed energy station based on LSTM according to the present invention;

[0072] Figure 5 Schematic diagram of the output gate of a distributed energy station heating load prediction method based on LSTM in the present invention;

[0073] Figure 6 This is a schematic diagram of the results of an embodiment of a distributed energy station heating load prediction method based on LSTM in the present invention. DETAILED DESCRIPTION

[0074] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0075] Example: Figures 1-6 As shown, the present invention provides a distributed energy station heating load prediction method based on LSTM: the method comprises the following steps:

[0076] S1. Obtain historical data of a distributed energy station, perform data missing processing based on the historical data, and record the number of missing data processed;

[0077] S2. Build an isolation forest model based on historical data to identify abnormal data, perform data missing processing on the abnormal data, and record it in the missing processing quantity;

[0078] S3. For the processed data, a distributed energy station load forecasting model is constructed based on the long short-term memory network;

[0079] S4. Build a system evaluation model based on the real-time operation data of the distributed energy station load to realize the update and reuse of the distributed energy station load prediction model.

[0080] The historical data includes: site operation data, weather data, building data and time data;

[0081] The site operation data includes supply water temperature, return water temperature, instantaneous heat, instantaneous flow, accumulated heat, accumulated electricity, indoor temperature and target room temperature; the weather data includes temperature, humidity, rainfall, wind speed, wind direction, weather type and solar radiation; the building data includes floor, heating area, building insulation level, building type and heating type; the time data includes the hour of the day, the week and whether it is a holiday.

[0082] The data missing processing includes:

[0083] When historical data is missing, choose to abandon or linear interpolate the processing method, including:

[0084] If there are complete data records before and after the missing data point, the missing data is supplemented by linear interpolation. Linear interpolation estimates the value of the missing point by drawing a straight line based on the known data before and after the missing data point.

[0085] The drawing method includes:

[0086] For example, in n-dimensional data, there is data before the missing point. , data exists after the missing point , the time difference between the two existing data is , the time difference between the missing point and the data before the missing point is , then the missing point data is:

[0087]

[0088] If the data before and after the missing data point is incomplete, or multiple data points are missing consecutively, the missing data point will be discarded to ensure the authenticity and reliability of the data.

[0089] For data that is difficult to determine manually whether it is abnormal, the isolation forest algorithm is used to screen for anomalies. Isolation forest is an unsupervised anomaly detection method based on a random binary tree. The process of detecting abnormal heating data based on isolation forest is as follows:

[0090] Because differences in weather, building, and time data can lead to significant variations in heating operation data, anomalies require grouping. All heating operation data is divided by weather, building, and time data ranges, with similar data grouped together. Anomaly detection is performed on each group using the isolation forest algorithm to detect any anomalies in the operation data.

[0091] The grouping process specifically includes:

[0092] Weather data: Weather codes and outdoor temperature are used as grouping criteria. Weather codes reflect changes in rainfall, humidity, and solar radiation. Outdoor temperature is divided into groups of 5 degrees Celsius, with the endpoints before and after included. For example, outdoor temperature in the range of [0,5) degrees Celsius is grouped together, and [5,10) degrees Celsius is grouped together.

[0093] Building data: Since the heating area, building insulation level, building type, and heating type values ​​for the same site are equal, the number of floors is used as the grouping basis;

[0094] Time data: divided by the time range of the day, whether it is a weekend, and whether it is a holiday. The time range of the day is divided into "daytime working hours" and "nighttime off-duty time". The "daytime working hours" and "nighttime off-duty time" are set according to the specific conditions of different sites. The default is 7:00 am to 7:00 pm and 7:00 pm to 7:00 am, where the endpoints are included before but not after.

[0095] S5-1. For any group, randomly select N groups of heating operation data as subsamples and put them into the root node of an isolated tree;

[0096] S5-2. At each root node, randomly select a feature (such as water supply temperature) from the heating operation data and randomly select a value between the maximum and minimum values ​​of the feature as the split value. The data is divided into the left subtree or the right subtree based on this split value.

[0097] S5-3, randomly select one of the other data items of the heating operation data as a feature, and repeat step S5-2 until the data cannot be divided any further, resulting in an isolated tree;

[0098] S5-4. Repeat steps S5-1 to S5-3 to obtain M isolated trees to form an isolation forest, where M is the number of isolated trees obtained by applying the isolation forest algorithm to the data in one group;

[0099] S5-5. In the isolation forest, a threshold for the number of split points is set and any data is judged. During the isolation process, if the number of split points exceeds the threshold, it is judged as normal data; if the number of split points does not exceed the threshold, it is judged as abnormal data.

[0100] S5-6. Feedback the detected abnormal data for data missing processing.

[0101] Due to the time delay of heating data, the predicted load value will not immediately affect the room temperature after it is issued. Instead, it takes some time for the heat to be transferred to each room and affect the room temperature. However, due to differences in the heating area and number of floors of each distributed energy station, the speed of heat transfer varies, making it impossible to uniformly align the time. Therefore, when conducting load forecasts, it is necessary to not only refer to the current real-time heating data, but also comprehensively consider the heating data over the past period of time.

[0102] The long short-term memory network is used to realize the load forecast of distributed energy stations. By introducing three gate control units, namely input gate, forget gate and output gate, we can control which dimensions of the heating data at the previous moment are retained and which dimensions of the latest real-time data are used to update the information to obtain the final forecast result. The unit structure of the model at a certain moment is as follows: Figure 2 shown.

[0103] The distributed energy station load forecasting model includes:

[0104] Three gating units, namely the input gate, forget gate, and output gate, are introduced to control the dimensions of the heating data retained at the previous moment. LSTM input data includes: supply water temperature, return water temperature, instantaneous heat, instantaneous flow, indoor temperature, target room temperature, temperature, humidity, rainfall, wind speed, wind direction, weather type, solar radiation, floor, hour of the day, week number, and whether it is a holiday.

[0105] like Figure 3 As shown, the forget gate determines the data information that LSTM discards from the heating data at the previous moment, specifically including:

[0106] Set the current time to t, and the forget gate reads the output of the previous layer and current input real-time data , calculate the retention factor of the information at the previous moment :

[0107]

[0108] Where: is the forget gate weight matrix, is the forget gate bias, For the sigmoid function, Mapped to the interval [0,1]; where, Generated by the input at the previous moment and the retained information of the previous moment;

[0109] The cell state obtained from the information of the previous moment and Multiply, A value of 1 indicates that All information in the , A value of 0 means completely discarding Information in;

[0110] like Figure 4 As shown, the input gate is used to filter information and determine the new heating data input at the current moment to be used to update the final output data information, including a sigmoid layer and a tanh layer;

[0111] The sigmoid layer is based on the output of the previous layer. and current input real-time data Calculate the update factor of heating information at the new time :

[0112]

[0113] in: is the input gate sigmoid layer weight matrix, Bias for the input gate sigmoid layer;

[0114] The tanh layer calculates candidate cell state values , is added to the cell state at the current moment middle:

[0115]

[0116] in: is the input gate tanh layer weight matrix, is the input gate tanh layer bias, is the tanh activation function;

[0117] The current cell state It is equal to the sum of the product of the output value of the forget gate and the product of the two output values ​​of the input gate, that is:

[0118]

[0119] like Figure 5 As shown, the output gate is used to calculate the output of the current layer ;

[0120] The output gate first passes through a sigmoid layer to determine the cell state at the current moment Output data information:

[0121]

[0122] Where: is the output gate weight matrix, is the output gate bias;

[0123] The current cell state By processing with tanh, a value in the interval [-1,1] is obtained and it is combined with Multiply and finally determine the output :

[0124]

[0125] If the current layer is the last layer, then the It is the final prediction value of the distributed energy station heating load prediction model.

[0126] Also includes:

[0127] Set the prediction time span p and reference data time span k for the LSTM algorithm;

[0128] Assume that the current time is d, and use k data from d-k+1 to d to predict the energy station load value at the future p-th time;

[0129] When LSTM is required to provide the current load recommendation value: During the model training phase, set p = 0, and the training set consists of k data items. The target value of each data group is the load value corresponding to the kth data item. The weather data is the weather data at the corresponding moment of each data item. During the model inference phase, select k-1 historical data items and the current real-time data, combine them with the weather data at the corresponding moment, and apply the model to obtain the current load recommendation value.

[0130] When LSTM is needed to predict the load recommendation value at a future time, set the future time R. In the model training phase, set p = Rd. The training set consists of k data items. The target value of each group of data is the load value corresponding to the k+pth data item. The weather data is the weather data at the time + p corresponding to each data item. In the model inference phase, select k-1 historical data items and the current real-time data, combine them with the weather data at the time + p corresponding to each data item, and apply the model to predict the load recommendation value at the future time R.

[0131] In this embodiment, a heat load prediction model of a distributed energy station is selected, and the prediction results of 50 sets of data are compared with the actual data. The results are as follows: Figure 6 , where the gray solid line is the real data, the black dotted line is the prediction result, the x-axis is 50 sets of data, and the y-axis is the load prediction value;

[0132] Select MSE and MAE as evaluation indicators, and the calculation formulas of MSE and MAE are:

[0133]

[0134]

[0135] Where, is the load forecast value of the i-th group of data, is the true load value of the i-th group of data.

[0136] Before starting training, the training set and test set were divided into two groups in a ratio of 9:1. The MSE calculation result of the model test set was 0.08 and the MAE was 0.2. The error was small compared with historical experience, and the model prediction effect was ideal.

[0137] The system evaluation model includes:

[0138] After the model has been running for a while, it may encounter new weather data that differs significantly from the historical data used in previous training, or user needs may change. In order to improve the generalization of the model, it is necessary to update the model, that is, re-execute all previous steps to generate a new model. The system evaluation model includes:

[0139] Obtain the generated distributed energy station heating load forecasting model and the amount of missing data in the model construction process;

[0140] Define a model prediction tolerance threshold and perform an operational analysis on the generated distributed energy station heating load forecasting model. The operational analysis includes:

[0141] Starting from the current moment, obtain the absolute value of the difference between the actual result and the predicted result each time, and continue to accumulate it until the total absolute value of the difference exceeds the model prediction tolerance threshold for the first time. Record the total number of predicted data experienced from the current moment;

[0142] With the number of missing data as the horizontal axis and the total number of experienced forecast data as the vertical axis, a linear fit is performed on the coordinate points formed by the heating load forecasting model of all distributed energy stations to form a functional relationship between the number of missing data and the total number of experienced forecast data;

[0143] Obtain the number of missing data in the current distributed energy station heating load forecasting model, feedback the total number of forecast data experienced based on the functional relationship, mark the position, and when the model reaches the marked position, feedback is sent to the management port to remind the model to update.

[0144] When a new site is put into use, its building attribute data is compared with that of the site in the existing heating load prediction model, and the sites that belong to the same group with its building attributes are calculated and reused. After the new site has been put into use for some time, the model is updated according to the historical operating data of the new site to ensure that the average room temperature of the site reaches a stable state.

[0145] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A distributed energy station heating load forecasting method based on LSTM, characterized by: The method comprises the following steps: S1. Obtain historical data of a distributed energy station, perform data missing processing based on the historical data, and record the number of missing data processed; S2. Build an isolation forest model based on historical data to identify abnormal data, perform data missing processing on the abnormal data, and record it in the missing processing quantity; S3. For the processed data, a distributed energy station load forecasting model is constructed based on the long short-term memory network; S4. Build a system evaluation model based on the real-time operation data of distributed energy station loads to achieve the update and reuse of distributed energy station load forecasting models; Set the prediction time span p and reference data time span k for the LSTM algorithm; Assume that the current time is d, and use k data from d-k+1 to d to predict the energy station load value at the future p-th time; When LSTM is required to provide the current load recommendation value: During the model training phase, set p = 0, and the training set consists of k data items. The target value of each data group is the load value corresponding to the kth data item. The weather data is the weather data at the corresponding moment of each data item. During the model inference phase, select k-1 historical data items and the current real-time data, combine them with the weather data at the corresponding moment, and apply the model to obtain the current load recommendation value. When LSTM is needed to predict the load recommendation value at a future time, set the future time R. During the model training phase, set p = Rd. The training set consists of k data items. The target value of each data group is the load value corresponding to the k+pth data item. The weather data is the weather data at the time + p corresponding to each data item. During the model inference phase, select k-1 historical data items and the current real-time data. Combined with the weather data at the time + p corresponding to each data item, the model is applied to predict the load recommendation value at the future time R. The system evaluation model includes: Obtain the generated distributed energy station heating load forecasting model and the amount of missing data in the model construction process; Define a model prediction tolerance threshold and perform an operational analysis on the generated distributed energy station heating load forecasting model. The operational analysis includes: Starting from the current moment, obtain the absolute value of the difference between the actual result and the predicted result each time, and continue to accumulate it until the total absolute value of the difference exceeds the model prediction tolerance threshold for the first time. Record the total number of predicted data experienced from the current moment; With the number of missing data as the horizontal axis and the total number of experienced forecast data as the vertical axis, a linear fit is performed on the coordinate points formed by the heating load forecasting model of all distributed energy stations to form a functional relationship between the number of missing data and the total number of experienced forecast data; Obtain the number of missing data in the current distributed energy station heating load forecasting model, feedback the total number of forecast data experienced based on the functional relationship, mark the position, and when the model reaches the marked position, feedback is sent to the management port to remind the model to update.

2. The LSTM-based distributed energy station heating load forecasting method according to claim 1 is characterized in that: The historical data includes: site operation data, weather data, building data and time data; The site operation data includes supply water temperature, return water temperature, instantaneous heat, instantaneous flow, accumulated heat, accumulated electricity, indoor temperature and target room temperature; the weather data includes temperature, humidity, rainfall, wind speed, wind direction, weather type and solar radiation; the building data includes floor, heating area, building insulation level, building type and heating type; the time data includes the hour of the day, the week and whether it is a holiday.

3. The LSTM-based distributed energy station heating load prediction method according to claim 1 is characterized in that: The data missing processing includes: When historical data is missing, choose to abandon or linear interpolate the processing method, including: If there are complete data records before and after the missing data point, the missing data is supplemented by linear interpolation. Linear interpolation estimates the value of the missing point by drawing a straight line based on the known data before and after the missing data point. Drawing methods include: Data exists before the missing point , data exists after the missing point , the time difference between the two existing data is , the time difference between the missing point and the data before the missing point is , then the missing point data is: If the data before and after the missing data point are incomplete, or if multiple data points are missing consecutively, choose to discard the missing data point.

4. The LSTM-based distributed energy station heating load prediction method according to claim 3 is characterized by: The abnormal data determination includes: Grouping data into similar groups. The grouping process specifically includes: Weather data: Weather codes and outdoor temperature are used as grouping criteria. Weather codes reflect changes in rainfall, humidity, and solar radiation. Outdoor temperature is divided into groups of 5 degrees Celsius, with the endpoints including the end point and excluding the end point. Building data: take the number of floors as the grouping basis; Time data: divided by the time range of the day, whether it is a weekend, and whether it is a holiday. The time range of the day is divided into daytime working hours and nighttime off-duty hours. The daytime working hours and nighttime off-duty hours are set according to the specific conditions of different sites. The default is 7:00 am to 7:00 pm and 7:00 pm to 7:00 am, where the endpoints are included before but not after.

5. The LSTM-based distributed energy station heating load prediction method according to claim 4 is characterized in that: Also includes: S5-1. For any group, randomly select N groups of heating operation data as subsamples and put them into the root node of an isolated tree; S5-2. At each root node, randomly select a feature from the heating operation data and randomly select a value between the maximum and minimum values ​​of the feature as the split value. The data is divided into the left subtree or the right subtree according to this split value. S5-3, randomly select one of the other data items of the heating operation data as a feature, and repeat step S5-2 until the data cannot be divided any further, resulting in an isolated tree; S5-4. Repeat steps S5-1 to S5-3 to obtain M isolated trees to form an isolation forest, where M is the number of isolated trees obtained by applying the isolation forest algorithm to the data in one group; S5-5. In the isolation forest, a threshold for the number of split points is set and any data is judged. During the isolation process, if the number of split points exceeds the threshold, it is judged as normal data; if the number of split points does not exceed the threshold, it is judged as abnormal data. S5-6. Feedback the detected abnormal data for data missing processing.

6. The LSTM-based distributed energy station heating load forecasting method according to claim 5, characterized in that: The distributed energy station load forecasting model includes: Three gating units, namely the input gate, forget gate, and output gate, are introduced to control the dimensions of the heating data retained at the previous moment. LSTM input data includes: supply water temperature, return water temperature, instantaneous heat, instantaneous flow, indoor temperature, target room temperature, temperature, humidity, rainfall, wind speed, wind direction, weather type, solar radiation, floor, hour of the day, week number, and whether it is a holiday. The forget gate determines the data information that LSTM discards from the heating data at the previous moment, specifically including: Set the current time to t, and the forget gate reads the output of the previous layer and current input real-time data , calculate the retention factor of the information at the previous moment : Where: is the forget gate weight matrix, is the forget gate bias, For the sigmoid function, Mapped to the interval [0,1]; where, Generated by the input at the previous moment and the retained information of the previous moment; The cell state obtained from the information of the previous moment and Multiply, A value of 1 indicates that All information in the , A value of 0 means completely discarding Information in; The input gate is used to filter information and determine the new heating data input at the current moment to be used to update the final output data information, including a sigmoid layer and a tanh layer; The sigmoid layer is based on the output of the previous layer. and current input real-time data Calculate the update factor of heating information at the new time : in: is the input gate sigmoid layer weight matrix, Bias for the input gate sigmoid layer; The tanh layer calculates candidate cell state values , is added to the cell state at the current moment middle: in: is the input gate tanh layer weight matrix, is the input gate tanh layer bias, is the tanh activation function; The current cell state It is equal to the sum of the product of the output value of the forget gate and the product of the two output values ​​of the input gate, that is: The output gate is used to calculate the output of the current layer ; The output gate first passes through a sigmoid layer to determine the cell state at the current moment Output data information: Where: is the output gate weight matrix, is the output gate bias; The current cell state By processing with tanh, a value in the interval [-1,1] is obtained and it is combined with Multiply and finally determine the output : If the current layer is the last layer, then the It is the final prediction value of the distributed energy station heating load prediction model.

7. The LSTM-based distributed energy station heating load prediction method according to claim 6, characterized in that: Also includes: When a new site is put into use, its building attribute data is compared with that of the site in the existing heating load prediction model, and the sites that belong to the same group with its building attributes are calculated and reused. After the new site has been put into use for some time, the model is updated according to the historical operating data of the new site to ensure that the average room temperature of the site reaches a stable state.

Images