An indoor temperature prediction apparatus, method, computer program performing the method

Abstract

The present invention relates to an artificial intelligence-based indoor temperature prediction device, an indoor temperature prediction method using the same, and a computer program for performing the method. The present invention discloses an artificial intelligence-based indoor temperature prediction device (100), comprising: a temperature prediction model generation unit (110) that generates a plurality of artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data; and an optimal prediction model derivation unit (120) that derives an optimal prediction model among the plurality of temperature prediction models.

Description

Technology Field

[0001] The present invention relates to an artificial intelligence-based indoor temperature prediction device, an indoor temperature prediction method using the same, and a computer program for performing the method. Background Technology

[0002] Since the Industrial Revolution, energy sources such as fossil fuels and natural gas have been the primary suppliers of heating and cooling energy. However, with the emergence of environmental issues in modern times, efforts are being made to find more eco-friendly energy sources or to explore technologies that can maximize the efficiency of heating and cooling energy.

[0003] To minimize a building's energy consumption, accurate prediction of the indoor temperature during use is required, which enables the achievement of both objectives: minimizing energy consumption and preserving indoor comfort.

[0004] There are attempts to apply machine learning or deep learning-based artificial intelligence technologies to predict indoor temperature, but there is a great need for improvement in terms of prediction accuracy and time required. The problem to be solved

[0005] The objective of the present invention is to provide an artificial intelligence-based indoor temperature prediction device capable of accurately and quickly predicting indoor temperature, an indoor temperature prediction method using the same, and a computer program for performing the method. means of solving the problem

[0006] The present invention, created to achieve the purpose of the present invention as described above, discloses an artificial intelligence-based indoor temperature prediction device (100), comprising: a temperature prediction model generation unit (110) that generates a plurality of artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data; and an optimal prediction model derivation unit (120) that derives an optimal prediction model among the plurality of temperature prediction models.

[0007] The above indoor temperature prediction device (100) may further include an indoor temperature prediction unit (130) that predicts the indoor temperature (Tr) using the above optimal prediction model.

[0008] The temperature prediction model generation unit (110) can generate the plurality of temperature prediction models using a deep learning algorithm based on an Artificial Neural Network (ANN).

[0009] The above optimal prediction model derivation unit (120) can derive the temperature prediction model closest to the average of the plurality of temperature prediction curves derived from the plurality of temperature prediction models among the plurality of temperature prediction models as the optimal prediction model.

[0010] The above optimal prediction model derivation unit (120) derives a past synchronization section (PM) in which the distribution of the ambient temperature (To) is similar, and based on the correlation (C_pm(Tr_f, Tr_a)) and similarity (rmse_pm(Tr_f, Tr_a)) between the temperature prediction value (Tr_f) and the actual temperature value (Tr_a) in the above synchronization section (PM), one of the above multiple temperature prediction models can be derived as the optimal prediction model.

[0011] The optimal prediction model derivation unit (120) can derive the optimal prediction model by calculating the synchronization index (Ipm) using the correlation (C(Tr_f, Tr_a)) and similarity (rmse(Tr_f, Tr_a)) and selecting a preset number of temperature prediction models based on the synchronization index (Ipm).

[0012] The above optimal prediction model derivation unit (120) can derive the temperature prediction model closest to the average of the temperature prediction curves derived from the above preset number of selected temperature prediction models as the optimal prediction model.

[0013] The optimal prediction model derivation unit (120) can derive the optimal prediction model by using the average temperature probability distribution (P(Tave)) for the average temperature prediction value (Tave) of the temperature prediction values ​​(Tr_f) derived from each of the plurality of temperature prediction models, and the correlation (C(To, Tr_f)) between the ambient temperature (To) and the temperature prediction value (Tr_f).

[0014] The optimal prediction model derivation unit (120) can derive the optimal prediction model by calculating a probability index (Ipb) that corrects the correlation (C(To, Tr_f) to the Gaussian distribution of the average temperature probability distribution (P(Tave)), and by selecting a preset number of temperature prediction models based on the probability index (Ipd) among the plurality of temperature prediction models.

[0015] The above optimal prediction model derivation unit (120) can derive the temperature prediction model closest to the average of the temperature prediction curves derived from the above preset number of selected temperature prediction models as the optimal prediction model.

[0016] In another aspect, the present invention discloses an indoor temperature prediction method performed in an artificial intelligence-based indoor temperature prediction device (100), characterized by comprising: a temperature prediction model generation step of generating a plurality of artificial intelligence-based temperature prediction models that generate a temperature prediction curve of the indoor temperature by learning a preset dataset as training data; and an optimal prediction model derivation step of deriving an optimal prediction model among the plurality of temperature prediction models.

[0017] The above indoor temperature prediction method may further include an indoor temperature prediction step of predicting the indoor temperature (Tr) using the above optimal prediction model.

[0018] In another aspect, the present invention discloses a computer program stored on a computer-readable recording medium for performing a method for predicting indoor temperature. Effects of the invention

[0019] The artificial intelligence-based indoor temperature prediction device according to the present invention, the indoor temperature prediction method using the same, and the computer program for performing the method have the advantage of being able to predict the indoor temperature accurately and quickly. Brief explanation of the drawing

[0020] FIG. 1 is a block diagram illustrating an indoor temperature prediction device according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating an indoor temperature prediction method performed in an indoor temperature prediction device according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating an indoor temperature prediction method performed in an indoor temperature prediction device according to one embodiment of the present invention. FIGS. 3a and FIGS. 3b are drawings illustrating one example of a building that is subject to indoor temperature prediction according to one embodiment of the present invention. Figures 4a and 4b are graphs illustrating examples of the occupancy rate and the usage rate of lighting fixtures / equipment by day of the week and time of day. FIG. 5 is a conceptual diagram illustrating a system for simulating heating and cooling control within the building of FIG. 3a and FIG. 3b. Figure 6 is a graph showing the hourly heating and cooling loads derived through the simulation system of Figure 5. Figure 7a is a conceptual diagram showing the error between the actual value and the predicted value when the indoor temperature is predicted in the short term, and Figure 7b is a conceptual diagram showing the error between the actual value and the predicted value when the indoor temperature is predicted in the long term. Figures 8a and 8b are graphs showing the long-term prediction results of indoor temperature derived using an artificial intelligence-based prediction model. Figure 9 is a graph showing the long-term prediction results of indoor temperature derived according to an artificial intelligence-based indoor temperature prediction model. Figure 10 is a graph showing the verification results of a short-term prediction using the artificial intelligence-based indoor temperature prediction model of Figure 9. Figure 11 is a graph showing the long-term indoor temperature prediction results based on artificial intelligence. FIG. 12 is a graph showing the long-term indoor temperature prediction results according to the indoor temperature prediction method according to one embodiment of the present invention. FIGS. 13a to 13e are graphs for explaining an indoor temperature prediction method according to another embodiment of the present invention. FIG. 14 is a graph showing the long-term indoor temperature prediction results derived by the indoor temperature prediction method according to another embodiment of the present invention. FIG. 15 is a graph comparing long-term indoor temperature prediction results derived by an indoor temperature prediction method according to various embodiments of the present invention. Specific details for implementing the invention

[0021] The following describes an artificial intelligence-based indoor temperature prediction method, an apparatus for performing the same, and a computer program according to the present invention with reference to the attached drawings.

[0022] The artificial intelligence-based indoor temperature prediction device according to the present invention is a computing device and can communicate by being connected to one or more other servers or terminals through a network.

[0023] That is, the above indoor temperature prediction device corresponds to a computing device connected via a network with other servers, terminals, DBs, etc., and can be implemented as, for example, a server, desktop, laptop, tablet PC, or smartphone, and may include a communication network module (140) for wired / wireless network connection, a user input / output interface module (150) for user input / output, and a memory (storage device, 160).

[0024] The above indoor temperature prediction device includes a processor for performing the indoor temperature prediction method described below, and as illustrated in FIG. 1, may include a temperature prediction model generation unit (110) that generates a plurality of artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data, and an optimal prediction model derivation unit (120) that derives an optimal prediction model among the plurality of temperature prediction models.

[0025] The above temperature prediction model generation unit (110) is configured to generate multiple artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data, and various configurations are possible.

[0026] The above dataset may include one or more variables that enable the construction of a temperature prediction model to output a temperature prediction value (Tr_f) of the indoor temperature (Tr).

[0027] The above variables may consist of data collected over a preset period, and the data collection method may be implemented in various ways. For example, the above dataset may be collected through weather data provided by the Korea Meteorological Administration and programs that provide simulations of heat loads in buildings / indoors.

[0028] That is, some of the data included in the dataset for constructing the indoor temperature prediction model according to the present invention may be obtained through a simulation program or by measurement, but the scope of the present invention is not limited thereto, and it is obvious that the dataset can be collected and utilized in various ways.

[0029] Below, the above simulation program is a program capable of calculating energy consumption efficiency by predicting the power consumption of a building / indoor, and an exemplary method of obtaining a dataset through TRNSYS is described in detail.

[0030] This simulation program enables obtaining data on the heating and cooling load of a building, and can define a building to be simulated as shown in FIGS. 3a and 3b.

[0031] Referring to FIGS. 3a and 3b, the subject building has walls facing east (E), west (W), south (S), and north (N), and a glass door (G) may be installed on the east wall, an entrance door (D) on the west wall, and a window (WD) on the north wall. Additionally, it may include a space where heating and cooling are performed directly, and a toilet (T) into which a certain flow rate is introduced without direct heating or cooling.

[0032] In addition, to account for the thermal effects caused by indoor occupants, the number of indoor occupants, occupancy rates by time of day, thermal effects per person, load per area due to lighting and equipment, and usage rates by time of day can be considered.

[0033] The operation of the heating and cooling unit can be configured to be controlled according to the outside temperature (To), temperature setpoint, occupancy status, etc. For example, the ratio of material occupancy according to time and day of the week, and the ratio of lighting / device usage according to time and day of the week, can be set as shown in FIG. 4a and FIG. 4b. These setting values ​​can be adjusted.

[0034] Figure 5 is a schematic diagram illustrating the operation of the TRNSYS program, which can be configured to perform energy analysis using meteorological data from a specific region (e.g., Daegu Metropolitan City) during a specific period (e.g., 2021).

[0035] Here, the setting temperature of the air conditioner and heater can also be set. For example, when there are occupants, the air conditioner can be set to 25°C and the heater to 21°C, and when there are no occupants, the air conditioner can be set to 27°C and the heater to 19°C.

[0036] That is, for winter heating, the heater can be set to turn on when the temperature drops 1°C below the set temperature (19°C or 21°C) and turn off when the temperature rises 1°C above the set temperature (19°C or 21°C) (dead band 2°C).

[0037] Similarly, for summer cooling, the cooling unit can be set to turn on when the temperature rises 1°C above the set temperature (25°C or 27°C) and when the temperature drops 1°C below the set temperature (25°C or 27°C) (dead band 2°C).

[0038] Meanwhile, meteorological data required for simulation calculations, such as temperature (°C), humidity (%), wind speed (m / s), wind direction, local pressure (hpa), elevation angle, azimuth angle, ground reference, dew point temperature (°C), and solar radiation (MJ / m2), can be imported into the program from the Korea Meteorological Administration Data Open Portal (kma.go.kr) and the Astronomy and Space Science Information Portal (astro.kasi.re.kr).

[0039] A dataset for a specific period can be created using the results derived from the simulation program described above. For example, a dataset for two years can be created, but it is obvious that the period can be set in various ways. An indoor temperature prediction model can be trained using a portion of the obtained dataset and validated using the remaining portion to predict the indoor temperature (Tr).

[0040] However, the TRNSYS program operation considers 0:00 on January 1st as the start time of Monday and processes the day of the week change at 24-hour intervals.

[0041] If a specific day within this year is referred to as "today," today's data can be predicted using an AI prediction model based on the dataset up to yesterday. In this case, since today's outside temperature is data that can be known in advance through the Korea Meteorological Administration's weather forecast, it can be understood as data usable for predicting today's data.

[0042] A dataset usable as training data for constructing the artificial intelligence-based indoor temperature prediction model of the present invention may include, for example, the following variables.

[0043] - Hour

[0044] - Tr: Indoor temperature

[0045] - Cooling Load (Total_Cooling(KJ / hr))

[0046] - Heating Load (Total_Heating(KJ / hr))

[0047] - Heat pump load (HP_Energe_input_total(KJ / hr)): Energy used by the entire heat pump system

[0048] - Heating setting temperature (T_Heat_setting)

[0049] - Cooling setting temperature (T_Cool_setting)

[0050] - HVAC outlet temperature (T_HVACoutlet)

[0051] - HVAC outlet flow rate (Flowrate_HVACoutlet(kg / hr))

[0052] - Occupancy (Occupancy_office)

[0053] - Heating control signal (HeaterON)

[0054] - Cooling control signal (CoolerON)

[0055] - Sky temperature (°C): Sky temperature

[0056] - Ambient temperature (°C)

[0057] After obtaining the above dataset, an indoor temperature prediction model can be trained using a portion of the obtained dataset and validated using the remaining portion to predict the indoor temperature (Tr).

[0058] However, the term "prediction" here may encompass two meanings. The first is short-term forecasting, and the second is long-term forecasting. In this invention, short-term forecasting is distinguished as "prediction," and long-term forecasting as "forecasting."

[0059] Short-term prediction involves calculating only the value of the next step, namely 9:00 today, using actual data up to 8:55, which is immediately before the time to be predicted, when the time to be predicted is 9:00 today, and this can be schematically represented as in Fig. 7a.

[0060] Referring to Figure 7a, it can be seen that the difference between the prediction and the actual is not significant because the short-term prediction uses the actual value immediately before the step to be predicted.

[0061] On the other hand, long-term forecasting is given only the actual value at the beginning and predicts the next step using the predicted value calculated in the previous step. As illustrated in Fig. 7b, as the prediction step is repeated, the error between the actual and predicted values ​​may increase compared to short-term forecasting.

[0062] In the present invention, the forecast is intended to be a long-term forecast as shown in FIG. 7b, and the term forecast below may be understood to mean a long-term forecast unless otherwise stated.

[0063] Meanwhile, an indoor temperature prediction model that predicts the indoor temperature (Tr) within a building can be constructed using an artificial intelligence-based algorithm.

[0064] For example, it can be constructed by applying the Recurrent Neural Network (RNN) algorithm, particularly Long Short Term Memory (LSTM), because the Recurrent Neural Network (RNN) is known to process sequential or time-dependent information well compared to the Artificial Neural Network (ANN) because it is possible to remember past learning information and reflect it in new learning results.

[0065] Figures 8a and 8b are graphs showing the results of predicting changes in indoor temperature (Tr) in early May using a prediction model built based on an LSTM algorithm. The black curve represents the training interval for building the prediction model, where the actual indoor temperature value (Tr_a) is plotted, and the predicted indoor temperature value (Tr_f) in the purple curve section (prediction interval) is plotted as a red curve using the trained prediction model.

[0066] Even though both Fig. 8a and Fig. 8b were calculated under the same conditions, it can be seen that the two prediction results appear completely different due to the accumulation of inaccurate prediction results and the randomness inherent in the program.

[0067] The prediction accuracy of Figure 8a is 98.3%, the correlation coefficient between the actual temperature value (Tr_a) and the predicted temperature value (Tr_f) in the prediction interval is 0.084, and the root mean square error (rmse) is 0.55. The prediction accuracy of Figure 8b is 95.6%, the correlation coefficient between the actual temperature value (Tr_a) and the predicted temperature value (Tr_f) in the prediction interval is -0.056, and the root mean square error (rmse) is 1.24.

[0068] In this case, since the indoor temperature is controlled within the prediction interval, the difference between the actual temperature value (Tr_a) and the predicted temperature value (Tr_f) is small compared to the magnitude of the indoor temperature (Tr) itself, so the accuracy appears relatively high. However, referring to the graph, it can be seen that the predicted temperature value (Tr_f) does not follow the behavior of the actual indoor temperature (Tr) well. This can be confirmed from the fact that the correlation coefficient appears near 0.

[0069] In the calculations of Figures 8a and 8b, even though the sliding window was set to a long period of one day, the prediction results varied each time the calculation was performed, and it took approximately 340 minutes to perform one calculation even with a 12-core Intel i5 equipped with an Nvidia Geforce 3070. Due to this inaccuracy and long computation time, LSTM is difficult to regard as a suitable artificial intelligence algorithm, and an indoor temperature prediction model was intended to be built using a multi-layer perceptron (MLP), which is a method of ANN.

[0070] MLP is a machine learning algorithm that mimics the principles of human neural networks and is the basis of deep learning, a field of machine learning. It consists of an input layer that receives input data, an output layer that outputs output data, and a hidden layer between them, and finds the weight and bias values ​​at the connection points to best predict the data.

[0071] To account for the historical impact on the current data, a sliding window technique was used by adding a few historical data points.

[0072] For example, the time (hr) interval was set to 5 minutes, and the window size for additionally using previous data was set to 10. As a result, one sliding window corresponds to a time interval of 50 minutes.

[0073] In addition, a sliding window was applied to important data from the past year to scale the data so that the maximum value was 1. 90% of these datasets were used as training data, and the remaining 10% were used for validation. After dividing the datasets, the data was randomly shuffled within each region, and training and validation were performed using this.

[0074] The following parameters were selected by calculating accuracy during the training period while varying the MLP parameters. However, since this is for illustrative purposes only, the parameters below may be selected differently.

[0075] Batch size: 75,

[0076] Number of epochs: 200,

[0077] Number of hidden nodes: 100

[0078] Activation function: sigmoid

[0079] At this time, the input data used for the ANN was set as follows, but it is obvious that this is also for illustrative purposes.

[0080] INPUTs: (48)

[0081] Weekly time (weekly time data), Daily time (daily time data), Day of the week, Week number (which week it is), Outdoor temperature (To), Initial indoor temperature, Difference between indoor temperature (Tr) and outdoor temperature (To), Heating / cooling operation status ('onoff')

[0082] Here, 'onoff' is the operating status of the air conditioner / heater, where 0 means all off, 1 means air conditioner ON, and 2 means heater ON.

[0083] Among these data, the outdoor temperature (To), initial indoor temperature, the difference between the indoor temperature (Tr) and the outdoor temperature (To), and the heating and cooling operation status ('onoff') can include a total of 48 data points by applying 10 sliding windows to add 10 previous data points.

[0084] The output data (OUTPUT) can be one internal temperature (Tr).

[0085] The indoor temperature (Tr) may vary depending on various outdoor conditions, occupancy, load of internal equipment, and day of the week patterns.

[0086] When the indoor temperature (Tr) reaches the operating temperature of the heating and cooling unit, the thermostat sends an ON / OFF control signal to the unit, which in turn affects the indoor temperature (Tr). The thermostat used in the TRNSYS calculation is a simple type; since the control signal for the heating and cooling unit is determined once the internal temperature (Tr) is determined, this signal can be treated as a known value for convenience.

[0087] Meanwhile, Figure 9 illustrates the results of a long-term forecast of today's and tomorrow's indoor temperatures (Tr) using last year's data. It shows that while the daytime period, where the temperature is frequently controlled, is predicted to some extent, the evening period is not predicted well. Accordingly, the calculation was performed by dividing the day into two 12-hour intervals: evening to morning and morning to evening.

[0088] Next, Figure 10 shows the results of short-term prediction of indoor temperature (Tr) using an ANN algorithm-based prediction model with 0.9 years of data from a 1-year dataset, followed by validation of the remaining 0.1 years of data. The accuracy is 99.8%, indicating that weights that closely track the actual data were selected.

[0089] However, the calculation in Fig. 10 is at a given moment (t iFor the calculation of ), t0~t up to immediately before i-1 Up to this point, it corresponds to a short-term prediction calculated using actual values, and it is impossible to guarantee how the long-term prediction, which is the accumulation of short-term prediction results, will turn out.

[0090] In neural network circuits, calculation results vary slightly each time due to initial conditions, inherent randomness, and results that are not 100% accurate. In the case of short-term prediction, the difference is not very large, but in the case of long-term forecasting, where differences at every moment accumulate, the difference can become large.

[0091] Figure 11 shows 10 long-term prediction results for the evening to morning hours of two dates using weights found by learning with the same data and computational parameters.

[0092] In the graph on the left, the date around 6,954 hours is October 17 of last year, and the actual prediction interval corresponds to May 4 of this year, starting from 11,730 hours in the graph on the right. For reference, October 17 of last year was selected as the subject of comparison because the change in ambient temperature is most similar to the prediction interval of May 4 of this year.

[0093] Calculations were performed on the target prediction interval for May 4 of this year using weight values ​​calculated for the data from October 17 of last year, and indoor temperature prediction models with the same weight values ​​for the left and right graphs were displayed in the same color.

[0094] Referring to Figure 11, two points can be observed. First, the temperature prediction value (Tr_f), which is the output result from all indoor temperature prediction models, did not predict the actual temperature value (Tr_a), indicated by the red curve, in the same way. This is due to the aforementioned inherent randomness and the accumulation of errors. Second, even if an indoor temperature prediction model is used where the past temperature prediction value (Tr_f) and the actual temperature value (Tr_a) are identical, it does not necessarily mean that the indoor temperature (Tr) is accurately predicted within the prediction interval. In other words, even if the prediction model indicated by the blue curve accurately predicted the indoor temperature (Tr) in the past, specifically in October of last year, it cannot be concluded that the same prediction model will accurately predict it in May of this year.

[0095] In other words, in the case of long-term forecasting, it can be reduced to a choice of which of the various indoor temperature prediction models calculated by MLPs best predicts the indoor temperature (Tr).

[0096] To this end, the temperature prediction model generation unit (110) can generate the plurality of temperature prediction models using an ANN (Artificial Neural Network)-based deep learning algorithm.

[0097] As explained in detail above, the plurality of temperature prediction models are all generated using an ANN algorithm with the same parameters for the same dataset, but they can all output different temperature prediction curves for indoor temperature (Tr).

[0098] The above temperature prediction curve is a curve that outputs the indoor temperature prediction value (Tr_f) according to time (hr) and can be constructed to accurately track the actual indoor temperature value (Tr_a).

[0099] The number of temperature prediction models generated by the temperature prediction model generation unit (110) may vary, but the following example describes a case where 300 temperature prediction models are generated by the temperature prediction model generation unit (110).

[0100] The above optimal prediction model derivation unit (120) can be configured in various ways to derive the optimal prediction model among a plurality of generated temperature prediction models.

[0101] Specifically, the optimal prediction model derivation unit (120) can select one temperature prediction model among a plurality of temperature prediction models that is estimated to best track the actual indoor temperature value (Tr_a), and the selected temperature prediction model can be defined as the optimal prediction model.

[0102] As a first embodiment for deriving the optimal prediction model, the optimal prediction model derivation unit (120) can derive the temperature prediction model closest to the average of the plurality of temperature prediction curves derived from the plurality of temperature prediction models among the plurality of temperature prediction models as the optimal prediction model.

[0103] For example, when 300 temperature prediction models are generated, the optimal prediction model derivation unit (120) can calculate the time average of the temperature prediction curves of the 300 temperature prediction models. Among the 300 temperature prediction models, a temperature prediction model having a temperature prediction curve closest to the calculated time average can be derived, and the derived temperature prediction model can be the optimal prediction model.

[0104] Referring to Figure 12, the prediction curve of the optimal prediction model derived from the average of 300 total temperature prediction models is plotted as the green curve (long-term prediction (overall average)), and it can be seen that the temperature prediction value (Tr_f) of the prediction interval shows a result that is very close to the actual temperature value (Tr_a) (red curve).

[0105] As a second embodiment for deriving the above optimal prediction model, the optimal prediction model derivation unit (120) can derive the optimal prediction model by using a past synchronization period (PM) in which the distribution of the outside temperature (To) is similar.

[0106] The above-mentioned synchronization interval (PM) is a past interval in which the distribution of the ambient temperature (To) is similar to the distribution of the ambient temperature (To) of the prediction interval to be predicted, and can be understood as an interval that acts as a pacemaker to improve the prediction accuracy of the prediction interval.

[0107] The basic idea of ​​the second embodiment is that since the outdoor temperature (To) is the factor that has the greatest influence on the indoor temperature (Tr), similar outdoor temperatures (To) will result in similar indoor temperature (Tr) patterns, and weekends and weekdays, where occupant patterns differ, can be considered separately.

[0108] The distribution of the outdoor temperature (To) and the initial indoor temperature (Tr_initial) in the above synchronization interval (PM) may be identical or similar to the distribution of the outdoor temperature (To) and the initial indoor temperature (Tr_initial) in the prediction interval. It is desirable that the difference between the initial indoor temperature (Tr_initial) in the synchronization interval (PM) and the prediction interval be 0.5℃ or less.

[0109] That is, the optimal prediction model derivation unit (120) selects a past date range having an outdoor temperature (To) distribution most similar to the outdoor temperature (To) distribution of today (prediction range) forecasted through the Korea Meteorological Administration as a synchronization range (PM, pacemaker), and can predict the indoor temperature (Tr) of the prediction range (today) by using a temperature prediction model that best matches the indoor temperature (Tr) of this synchronization range (PM, pacemaker). The curve on the left side of FIGS. 11 and FIGS. 12 illustrates the synchronization range (PM) selected based on these criteria.

[0110] First, the optimal prediction model derivation unit (120) can select one or more pre-set number of temperature prediction models that accurately match the indoor temperature (Tr) in the synchronization interval (PM).

[0111] For example, the optimal prediction model derivation unit (120) can select 5% to 15% of the total number of temperature prediction models, and more preferably, can select 9% to 11% of the total number of temperature prediction models.

[0112] The criteria for selecting a temperature prediction model that accurately predicts the indoor temperature (Tr) within the synchronization interval (PM) among all temperature prediction models can be constructed in various ways, for example, based on the correlation (C_pm(Tr_f, Tr_a)) and similarity (rmse_pm(Tr_f, Tr_a)) between the temperature prediction value (Tr_f) and the actual temperature value (Tr_a) within the aforementioned synchronization interval (PM).

[0113] The optimal prediction model derivation unit (120) calculates a synchronization index (Ipm) using correlation (C(Tr_f, Tr_a)) and similarity (rmse(Tr_f, Tr_a)) in a synchronization interval (PM), and can select a preset number of temperature prediction models based on the synchronization index (Ipm). The temperature prediction models selected based on the synchronization index (Ipm) can be understood as prediction models that output an optimal temperature prediction curve in the synchronization interval (PM).

[0114] Specifically, the above-mentioned synchronization index (Ipm) can be calculated using the following formula (1).

[0115] (1)

[0116] The above correlation (C(Tr_f, Tr_a)) indicates the trend of the temperature prediction value (Tr_f) and the actual temperature value (Tr_a) in the synchronization interval (PM). It means that the closer it is to 1, the more similar the trend, and the closer it is to 0, the less correlation there is between them.

[0117] In addition, the above similarity (rmse(Tr_f, Tr_a)) is a squared mean error value showing the difference between the temperature prediction value (Tr_f) and the actual temperature value (Tr_a), and the smaller the value, the more similar the two values ​​are.

[0118] Therefore, since the temperature prediction curve with a smaller value of the above synchronization index (Ipm) follows the actual indoor temperature curve in the synchronization interval (PM) better, the above optimal prediction model derivation unit (120) can select 30 temperature prediction models out of a number of preset numbers, for example, 300 total temperature prediction models, in order of smallest synchronization index (Ipm).

[0119] When a preset number of temperature prediction models (e.g., 30) are selected, the optimal prediction model derivation unit (120) can calculate the time average of the selected temperature prediction curves.

[0120] The above optimal prediction model derivation unit (120) can derive the temperature prediction model closest to the average of the temperature prediction curves derived from the above preset number of selected temperature prediction models as the optimal prediction model.

[0121] That is, when there are a total of 300 temperature prediction models, the optimal prediction model derivation unit (120) can derive the temperature prediction curve closest to the time average calculated from 30 temperature prediction models selected based on the synchronization index (Ipm) in the synchronization interval (PM) among the total 300 temperature prediction models. The derived temperature prediction model can be the optimal prediction model.

[0122] Referring to Figure 12, the time-averaged values ​​of 30 temperature prediction models selected in the synchronization interval (PM) based on the synchronization index (Ipm) were plotted as the blue curve (long-term prediction (average of 30)). For the blue curve, the correlation coefficient in the synchronization interval (PM) was 0.999, the similarity (rmse) was 0.139, and the accuracy was 0.994, while for the prediction interval, the correlation coefficient was 0.996, the similarity (rmes) was 0.363, and the accuracy was 0.9994.

[0123] The temperature prediction model closest to the blue curve representing the average of the 30 selected temperature prediction models among all temperature prediction models can be derived as the final optimal prediction model, and this optimal prediction model is illustrated as the purple curve (long-term prediction (optimization model)) in Fig. 12.

[0124] It can be seen that the temperature prediction value (Tr_f) of the prediction interval of the prediction curve (purple curve (long-term prediction (optimization model)) of the optimal prediction model derived through the second embodiment shows a result that is very close to the actual temperature value (Tr_a) (red curve).

[0125] As a third embodiment for deriving the above optimal prediction model, the optimal prediction model derivation unit (120) can derive an optimal prediction model capable of accurately predicting the indoor temperature (Tr) of the prediction interval by using the average temperature probability distribution of a plurality of generated temperature prediction models.

[0126] The basic idea of ​​the third embodiment is to examine temperature prediction models that have performed well in long-term forecasting using ANNs and to identify the accurate prediction value based on their characteristics. Since today's actual indoor temperature (Tr_a) cannot be known in advance, the error between today's actual indoor temperature (Tr_a) and the temperature prediction value (Tr_f) cannot be calculated; however, by examining various factors that influence this error, the value at which this error is minimized can be found.

[0127] First, Figure 13a shows the rmse (mean squared error) between the predicted temperature value (Tr_f) and the actual temperature value (Tr_a) for each temperature prediction model for the synchronization period (PM) and the prediction period (today). The horizontal axis represents rmse_pm(Tr_f, Tr_a) in the synchronization period (PM), and the vertical axis represents rmse(Tr_f, Tr_a) in the prediction period (today).

[0128] Since the past period with an ambient temperature (To) distribution similar to the forecast period (today) was selected as the synchronization period (PM), it can be seen that as the error in the synchronization period (PM) increases, the error in the forecast period (today) also generally increases. However, it can be seen that there is no exact linear relationship between them, and that even if the error in the synchronization period (PM) is minimized, it is not guaranteed to be minimized in the forecast period (today).

[0129] Accordingly, the optimal prediction model derivation unit (120) can derive the optimal prediction model using the average temperature probability distribution (P(Tave)) for the average temperature prediction value (Tave) of the temperature prediction values ​​(Tr_f) derived from each of the plurality of temperature prediction models, and the correlation (C(To, Tr_f)) between the ambient temperature (To) and the temperature prediction value (Tr_f).

[0130] Figure 13b shows the relationship between the average value of the predicted temperature value (Tr_f) in the prediction interval (today) (i.e., average temperature prediction value (Tave)) and the mean squared error (rmse(Tr_f, Tr_a)) between the predicted temperature value (Tr_f) in the prediction interval (today) and the actual temperature value (Tr_a).

[0131] It can be observed that the average temperature prediction value (Tave)) is clustered in one place in the region where the mean squared error (rmse(Tr_f, Tr_a)) of the prediction interval (today) is small, and as the mean squared error (rmse(Tr_f, Tr_a)) increases, the average temperature prediction value (Tave)) deviates from a certain range. This means that whether a temperature prediction model tracks the actual temperature value (Tr_a) well is related to the probability distribution of the average temperature prediction value (Tave).

[0132] In this regard, FIG. 13c illustrates the average temperature probability distribution (P(Tave)) for the average temperature prediction value (Tave) of the temperature prediction values ​​(Tr_f) derived from each of the plurality of temperature prediction models. Since the average temperature probability distribution (P(Tave)) for the average temperature prediction value (Tave) is discretely distributed, the average temperature probability distribution (P(Tave)) can be fitted into a Gaussian distribution (GP(P(Tave))) through Gaussian fitting.

[0133] Next, Figure 13d illustrates the relationship between the rmse (mean squared error) between the predicted temperature value (Tr_f) and the actual temperature value (Tr_a) for the prediction interval (today), and the correlation (C(To, Tr_f)) between the ambient temperature (To) and the predicted temperature value (Tr_f) for the prediction interval (today). The horizontal axis represents the correlation (C(To, Tr_f)) between the ambient temperature (To) and the predicted temperature value (Tr_f) for the prediction interval (today), and the vertical axis represents the rmse(Tr_f, Tr_a) for the prediction interval (today).

[0134] Figure 13d shows the effect of the correlation (C(To, Tr_f)) between the predicted temperature value (Tr_f) and the external temperature (To) on the error between the predicted temperature value (Tr_f) and the actual temperature value (Tr_a) in the prediction interval, where the larger the correlation (C(To, Tr_f)), the smaller the error between the predicted temperature value (Tr_f) and the actual temperature value (Tr_a) in the prediction interval. However, it can also be confirmed that it cannot be concluded that the error between the predicted temperature value (Tr_f) and the actual temperature value (Tr_a) in the prediction interval is minimized when the correlation (C(To, Tr_f)) is at its maximum.

[0135] As a result, the optimal prediction model derivation unit (120) can derive a temperature prediction model with a small error in the prediction section by appropriately combining the correlation (C(To, Tr_f)) between the outdoor temperature (To) and the indoor temperature prediction value (Tr_f), the average temperature probability distribution (P(Tave)) for the average temperature prediction value (Tave) of the predicted temperature prediction value (Tr_f), and the correlation (rmse_pm(Tr_f, Tr_a)) in the past synchronization section (PM).

[0136] First, the optimal prediction model derivation unit (120) can, as an optional configuration, first select some temperature prediction models based on the correlation (rmse_pm(Tr_f, Tr_a)) in the synchronization interval (PM) among all temperature prediction models.

[0137] For example, when a total of 300 temperature prediction models are generated, the optimal prediction model derivation unit (120) can first select 150 temperature prediction models that have a high correlation based on the correlation (rmse_pm(Tr_f, Tr_a)) in the synchronization interval (PM) among the 300 temperature prediction models.

[0138] Next, the optimal prediction model derivation unit (120) can correct the correlation (C(To, Tr_f)) between the ambient temperature (To) and the temperature prediction value (Tr_f) by using a Gaussian distribution (GP(P(Tave)) obtained by Gaussian fitting the average temperature probability distribution (P(Tave)) for the average temperature prediction value (Tave) of the temperature prediction value (Tr_f) predicted in the prediction interval (today).

[0139] The index obtained by correcting the above correlation (C(To, Tr_f)) by the Gaussian distribution (GP(P(Tave)) of the above average temperature probability distribution (P(Tave)) can be defined as the probability exponent (Ipb).

[0140] The relationship between the above probability index (Ipd) and the mean squared error (rmse(Tr_f, Tr_a)) of the prediction interval (today) is illustrated in Fig. 13e.

[0141] Referring to Fig. 13e, it can be seen that temperature prediction models with a large probability index (Ipd), which is a value corrected by the Gaussian distribution (GP(P(Tave))), correspond to temperature prediction models that are close to the minimum of the mean squared error (rmse(Tr_f, Tr_a)) of the prediction interval (today).

[0142] Accordingly, the optimal prediction model derivation unit (120) can derive the optimal prediction model by selecting a predetermined number of temperature prediction models based on the probability index (Ipd) among the plurality of temperature prediction models. For example, when 30 temperature prediction models are selected based on the probability index (Ipd), it cannot be concluded that the 30 temperature prediction models are the ones with the minimum squared mean error (rmse(Tr_f, Tr_a)) of the prediction interval (today), but it is clear that they are temperature prediction models with small errors.

[0143] Figure 14 shows the average curve of temperature prediction curves derived from 30 temperature prediction models selected based on the probability index (Ipd), plotted as a blue curve, and it can be seen that it follows the red curve representing the actual indoor temperature value (Tr_a) quite well.

[0144] Finally, the optimal prediction model derivation unit (120) can derive the temperature prediction model closest to the average of the temperature prediction curves derived from the temperature prediction models selected in the preset number as the optimal prediction model.

[0145] The final correlation coefficient obtained by the method according to the third embodiment was 0.995, the rmse was 0.162, and the accuracy was 0.992.

[0146] FIG. 15 illustrates the temperature prediction curve of the final optimal prediction model derived according to the first to third embodiments. It can be seen that through the method of the first to third embodiments, one of the total temperature prediction models (e.g., 300) can be selected as the optimal prediction model that closely tracks the actual temperature value (Tr_a).

[0147] At this time, the indoor temperature prediction device (100) may further include an indoor temperature prediction unit (130) that predicts the indoor temperature (Tr) of the prediction interval to be predicted using the derived optimal prediction model. It was confirmed that the indoor temperature prediction unit (130) accurately predicts the daily indoor temperature (Tr) for all seasons with an accuracy of 95% or higher in all methods according to the first to third embodiments.

[0148] In another aspect, the present invention may include an indoor temperature prediction method performed in an artificial intelligence-based indoor temperature prediction device (100), comprising: a temperature prediction model generation step of generating a plurality of artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data; and an optimal prediction model derivation step of deriving an optimal prediction model among the plurality of temperature prediction models.

[0149] The above indoor temperature prediction method may further include an indoor temperature prediction step that predicts the indoor temperature (Tr) using the derived optimal prediction model.

[0150] In another aspect, the present invention may be a computer program stored on a computer-readable recording medium to perform the indoor temperature prediction method described above.

[0152] The foregoing merely describes some preferred embodiments that can be implemented by the present invention. As is well known, the scope of the present invention should not be interpreted as being limited to the above embodiments, and all technical concepts that share the fundamental principles with the technical concept of the present invention described above shall be considered to be included within the scope of the present invention. Explanation of the symbols

[0153] 100: Indoor temperature prediction device

Claims

Claim 1 An artificial intelligence-based indoor temperature prediction device (100) comprises a temperature prediction model generation unit (110) that generates a plurality of artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data, and an optimal prediction model derivation unit (120) that derives an optimal prediction model among the plurality of temperature prediction models, wherein the optimal prediction model derivation unit (120) derives a past synchronization section (PM) in which the distribution of the outside temperature (To) is similar, and derives one of the plurality of temperature prediction models as the optimal prediction model based on the correlation (C_pm(Tr_f, Tr_a)) and similarity (rmse_pm(Tr_f, Tr_a)) between the temperature prediction value (Tr_f) and the actual temperature value (Tr_a) in the synchronization section (PM). Claim 2 An artificial intelligence-based indoor temperature prediction device (100) according to claim 1, further comprising an indoor temperature prediction unit (130) that predicts the indoor temperature (Tr) using the optimal prediction model. Claim 3 The artificial intelligence-based indoor temperature prediction device (100) according to claim 1, wherein the temperature prediction model generation unit (110) generates the plurality of temperature prediction models using an ANN (Artificial Neural Network)-based deep learning algorithm. Claim 4 An artificial intelligence-based indoor temperature prediction device (100) according to claim 1, wherein the optimal prediction model derivation unit (120) derives the temperature prediction model closest to the average of the plurality of temperature prediction curves derived from the plurality of temperature prediction models among the plurality of temperature prediction models as the optimal prediction model. Claim 5 delete Claim 6 In claim 1, the optimal prediction model derivation unit (120) calculates a synchronization index (Ipm) using correlation (C(Tr_f, Tr_a)) and similarity (rmse(Tr_f, Tr_a)), and selects a preset number of temperature prediction models based on the synchronization index (Ipm) to derive the optimal prediction model, thereby forming an artificial intelligence-based indoor temperature prediction device (100). Claim 7 In claim 6, the optimal prediction model derivation unit (120) derives the temperature prediction model closest to the average of the temperature prediction curves derived from the temperature prediction models selected in the preset number as the optimal prediction model, thereby forming an artificial intelligence-based indoor temperature prediction device (100). Claim 8 An artificial intelligence-based indoor temperature prediction device (100) comprises a temperature prediction model generation unit (110) that generates a plurality of artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data, and an optimal prediction model derivation unit (120) that derives an optimal prediction model among the plurality of temperature prediction models, wherein the optimal prediction model derivation unit (120) derives the optimal prediction model using an average temperature probability distribution (P(Tave)) for the average temperature prediction value (Tave) of the temperature prediction values ​​(Tr_f) derived from each of the plurality of temperature prediction models and a correlation (C(To, Tr_f)) between the outside temperature (To) and the temperature prediction value (Tr_f). Claim 9 In claim 8, the optimal prediction model derivation unit (120) calculates a probability index (Ipd) by correcting the correlation (C(To, Tr_f) to a Gaussian distribution of the average temperature probability distribution (P(Tave)), and selects a preset number of temperature prediction models based on the probability index (Ipd) among the plurality of temperature prediction models to derive the optimal prediction model, an artificial intelligence-based indoor temperature prediction device (100). Claim 10 In claim 9, the optimal prediction model derivation unit (120) derives the temperature prediction model closest to the average of the temperature prediction curves derived from the temperature prediction models selected in the preset number as the optimal prediction model, characterized in that it is an artificial intelligence-based indoor temperature prediction device (100). Claim 11 An indoor temperature prediction method performed in an artificial intelligence-based indoor temperature prediction device (100) according to any one of claims 1 and 8, comprising: a temperature prediction model generation step of generating a plurality of artificial intelligence-based temperature prediction models that output a temperature prediction curve of the indoor temperature by learning a preset dataset as training data; and an optimal prediction model derivation step of deriving an optimal prediction model among the plurality of temperature prediction models. Claim 12 A method for predicting indoor temperature according to claim 11, further comprising an indoor temperature prediction step of predicting the indoor temperature (Tr) using the optimal prediction model. Claim 13 A computer program stored on a computer-readable recording medium to perform the indoor temperature prediction method according to claim 12.

Images