Artificial intelligence-based power amount predictive artificial intelligence modeling training method using indoor and outdoor weather data

Abstract

A power amount predictive artificial intelligence modeling training method performed by a computing device comprising a processor, according to one embodiment of the present invention, may comprise the steps of: sensing indoor environment data of a specific household; receiving outdoor weather data of the region where the specific household is located; receiving power consumption data of the specific household; training an artificial intelligence modeling by using the indoor environment data, the outdoor weather data, and the power consumption data as parameters; calculating the degree of contribution of the parameters to power prediction accuracy on the basis of an actual power measurement value; and determining utilized parameters having a degree of contribution that is greater than or equal to a preset value among the parameters.

Description

AI Modeling Training Method for AI-Based Power Consumption Prediction Using Indoor and Outdoor Weather Data

[0001] The present invention relates to an artificial intelligence modeling learning method for predicting power consumption based on artificial intelligence using indoor and outdoor weather data.

[0002] Recently, IoT technology has been expanding across all industrial sectors. IoT stands for Internet of Things and refers to the technology of the Internet of Things. The Internet of Things is a technology that embeds sensors and communication functions into various objects to connect them to the Internet, meaning a technology that connects various objects via wireless communication. Here, "objects" can be various embedded systems, such as home appliances, mobile devices, and wearable devices. Objects connected to the Internet possess an identifier (ID) that distinguishes them and connect to the Internet; it is common for them to have embedded sensors to acquire data from the external environment. Each entity participating in the Internet of Things requires an ID that allows other entities to identify it. While RFID (Radio Frequency Identification) technology is used to represent the IDs of nearby objects, it is necessary to assign IP addresses to individual objects to verify their identities across a wider network.

[0003] As such, the market for IoT, a representative technology of the Fourth Industrial Revolution, is gradually growing, and as the number of IoT devices increases, it is becoming common practice to transmit values ​​acquired from sensors to databases or collection servers to analyze these data values.

[0004] Recently, as global carbon dioxide concentrations have been breaking all-time records every year (2022: 421 ppm), not only are abnormal climate phenomena occurring globally (such as the 49.6-degree heat dome in North America, Canada, and the once-in-a-thousand-year torrential rains in Zhengzhou, China), but there is also an urgent need to reduce greenhouse gas emissions to address these issues. In particular, in South Korea, the power energy sector has seen a dramatic surge in electricity consumption due to record-breaking heatwaves, while the supply reserve rate has hit a five-year low of 8.5%.

[0005] As an alternative, there is a growing demand for power management solutions targeting older homes and apartments, along with AI-based power consumption and reduction prediction technologies, in addition to efficient energy supply and demand through the establishment of smart grids.

[0006] Meanwhile, the aforementioned background technology is technical information that the inventor possessed for the derivation of the present invention or acquired during the process of deriving the present invention, and cannot necessarily be considered technology known to the general public prior to the filing of the present invention.

[0007] As mentioned above in the background technology section, there is a need to efficiently analyze Big Data, which consists of vast amounts of data; however, manually analyzing Big Data by humans is practically impossible and highly inefficient. To address this, artificial intelligence, particularly machine learning, can be utilized as an efficient model for analyzing Big Data. Generally, machine learning can proceed through the stages of data collection, data preprocessing, exploratory data analysis (EDA), model selection, and evaluation.

[0008] In particular, attempts are underway to predict power consumption using atmospheric conditions and outdoor weather data from various sensing devices. These power prediction technologies are classified into Long Short-Term Memory (LSTM) or Deep Neural Network (DNN) methods. While the LSTM method predicts power consumption based on past data, the DNN method improves accuracy by training a deep neural network with input data and correcting for incomplete results. Generally, the LSTM method can be used for academic papers or technical reports because power data is pre-established; however, for real-time web services, using LSTM presents significant difficulties in the event of other failures (such as system overload or failure to collect power data). In contrast, since DNNs predict consumption using pre-trained models (manual or automated learning), immediate response and action can be taken in the event of a failure.

[0009] The present invention aims to provide an artificial intelligence-based short-term power prediction method using indoor IoT atmospheric environment and outdoor weather data.

[0010] In addition, the present invention aims to process data for machine learning training purposes by performing data preprocessing to train sensing data collected from IoT using machine learning.

[0011] The above and other objects and advantages of the present invention will become apparent from the following description describing preferred embodiments.

[0012] A method for learning an artificial intelligence model for power consumption prediction, performed by a computing device including a processor according to an embodiment of the present invention, may include: sensing indoor environment data of a specific household; receiving outdoor weather data of a region where the specific household is located; receiving power consumption data of the specific household; learning an artificial intelligence model using indoor environment data, outdoor weather data, and power consumption data as parameters; calculating the contribution of the parameters to the accuracy of power prediction based on actual power values; and determining the parameters among the parameters that have a contribution greater than or equal to a preset value.

[0013] According to one embodiment, an artificial intelligence modeling learning method for power consumption prediction performs bias correction on a power prediction value (PRD) to provide a bias-corrected prediction result value (PRD biasCorr It may include additional steps to calculate ).

[0014] According to one embodiment, an artificial intelligence modeling learning method for power consumption prediction involves actual power values ​​(OBS) and bias-corrected predicted result values ​​(PRD biasCorr It may further include a step of verifying the bias-corrected prediction result value by calculating the Root Mean Squared Error (rMSE) using ).

[0015] According to one embodiment, indoor environment data includes sensing values ​​for temperature, humidity, fine dust, ultrafine dust, carbon monoxide, carbon dioxide, total volatile organic compounds, and formaldehyde, respectively, and weather data may include outdoor weather data, solar information data, indexed data, and time information data.

[0016] According to one embodiment, the indexed data includes a temperature-humidity index (THI) and a weather-competitive index (WCT), and the temperature-humidity index (THI) and the weather-competitive index (WCT) are each calculated using the following formula, and TA, HM, and WS may be temperature (°C), relative humidity (%), and wind speed (m / s), respectively.

[0017]

[0018] According to one embodiment, outdoor weather data includes total cloud cover, relative humidity, atmospheric pressure, temperature, dew point temperature, wind direction, wind speed, and solar radiation; solar information data includes solar zenith angle, solar azimuth angle, solar hour angle, atmospheric turbidity, total solar radiation on clear days, direct solar radiation on clear days, and scattered solar radiation on clear days; indexed data includes temperature-humidity index (THI) and weather perception index (WCT); and time information data may include information regarding whether it is a weekend, whether it is spring, whether it is summer, whether it is autumn, and whether it is winter.

[0019] According to one embodiment, a bias-corrected prediction result value (PRD biasCorr ) is calculated using the formula below, where a and b are arbitrary coefficients, and PRD may be a short-term forecast result.

[0020]

[0021] According to one embodiment, the root mean square error is calculated using the following formula, OBS is the actual power measurement value, and PRD biasCorr is the bias-corrected prediction result value, and n may be the number of bias correction processes.

[0022]

[0023] According to one embodiment, the preset value may be 20.

[0024] According to the present invention, when examining actual and predicted power consumption for a verification period for specific customers with regular / irregular patterns, it is possible to provide power prediction data similar to actual measurements at power peaks, as well as root mean square error and high correlation.

[0025] In addition, according to the present invention, the results of such customer location-based short-term power forecasts can be utilized as basic data for various innovative new services of the smart grid, and furthermore, not only can they contribute to housing-related carbon reduction policies through power consumption / savings per household, but they can also provide data based on participation in the Energy Pause National Demand Response (DR).

[0026] In addition, according to the present invention, energy saving and rationalization of use can be achieved by utilizing it as an energy management system for buildings as well as customer locations.

[0027] However, the effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below.

[0028] Figure 1 is a diagram briefly illustrating the basic concept of an artificial neural network (100) of machine learning.

[0029] FIG. 2a is a diagram briefly illustrating an indoor IoT system (200) according to one embodiment of the present invention.

[0030] FIG. 2b is a drawing showing an example of collecting outdoor weather information according to one embodiment of the present invention.

[0031] FIG. 3 is a diagram briefly illustrating a machine learning learning method (300) according to one embodiment of the present invention.

[0032] FIG. 4 is a diagram showing in detail the data preprocessing (320) process of machine learning according to one embodiment of the present invention.

[0033] FIG. 5 is a diagram showing a reference table of perceived temperature according to air temperature and wind speed according to one embodiment of the present invention.

[0034] FIG. 6 is a diagram showing the contribution according to input variables of manual learning modeling of short-term prediction according to one embodiment of the present invention.

[0035] FIG. 7 is a diagram showing the contribution according to input variables of an automatic learning model of short-term prediction according to one embodiment of the present invention.

[0036] FIG. 8 is a diagram illustrating an artificial intelligence modeling learning method for power consumption prediction according to an embodiment of the present invention.

[0037] The present invention will be described in detail below with reference to the embodiments and drawings. These embodiments are presented merely as examples to explain the invention more specifically, and it will be obvious to those skilled in the art that the scope of the invention is not limited by these embodiments.

[0038] Furthermore, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains, and in the event of a conflict, the description in this specification including definitions shall prevail.

[0039] To clearly explain the proposed invention in the drawings, parts unrelated to the description have been omitted, and similar parts throughout the specification have been given similar reference numerals. Furthermore, when a part is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Additionally, the term "part" as described in the specification refers to a single unit or block that performs a specific function.

[0040] In each step, identification codes (1st, 2nd, etc.) are used for convenience of explanation and do not describe the order of the steps; the steps may be performed differently from the specified order unless a specific order is clearly indicated in the context. That is, the steps may be performed in the same order as specified, substantially simultaneously, or in the reverse order.

[0041] Hereinafter, the structure of the present invention and the resulting effects are to be explained in more detail through specific embodiments and comparative examples. However, these embodiments are intended to explain the present invention more specifically, and the scope of the present invention is not limited to these embodiments.

[0042] Figure 1 is a diagram briefly illustrating the basic concept of an artificial neural network in machine learning.

[0043] As illustrated in FIG. 1, an artificial neural network (ANN, 100) may be configured as a layered structure including an input layer (110), an output layer (130), and at least one hidden layer (120-1, 120-2, 120-3) between the input layer (110) and the output layer (130). FIG. 1 illustrates an example having three hidden layers (120-1, 120-2, 120-3). Based on such a multi-layered structure, a deep learning algorithm can derive highly reliable results through learning that optimizes the weights of the activation functions between layers. Deep learning algorithms applicable to the present invention may include a deep neural network (DNN), a convolutional neural network (CNN), a recurrent neural network (RNN), etc. Deep neural networks (DNNs) are basically characterized by improving the results of learning by significantly increasing the number of intermediate layers (or hidden layers) within existing ANN models. For example, the above DNN is characterized by performing the learning process using two or more intermediate layers.

[0044] Accordingly, the computer can derive the optimal output value by repeating the process of generating classification labels, distorting the space, and distinguishing data.

[0045] Convolutional Neural Networks (CNNs) are characterized by a structure that extracts features from data and identifies patterns, unlike techniques that perform learning processes by extracting knowledge from existing data. The CNN described above can be executed through convolution and pooling processes. In other words, the CNN can include an algorithm composed of a combination of convolution layers and pooling layers. Here, the convolution layer performs the process of extracting data features (also known as the convolution process). This convolution process involves examining the adjacent components of each component in the data to identify features and deriving the identified features into a single image; as a compression process, it can effectively reduce the number of parameters. The pooling layer performs the process of reducing the size of the layers that have undergone the convolution process (also known as the pooling process). This pooling process can reduce the size of the data, cancel out noise, and provide consistent features in fine details. For example, the CNN described above can be utilized in various fields such as information extraction, sentence classification, and face recognition.

[0046] A recurrent neural network (RNN) is a type of artificial neural network specialized for learning repetitive and sequential data, characterized by having an internal recurrent structure. The RNN utilizes this recurrent structure to apply weights to past learning content and reflect them in current learning, thereby enabling a connection between current learning and past learning and possessing the characteristic of being time-dependent. As an algorithm that resolves the limitations of existing continuous, repetitive, and sequential data learning, the RNN can be utilized for tasks such as identifying speech waveforms or identifying the preceding and succeeding components of text. However, these are merely examples of specific deep learning techniques applicable to the present invention, and other deep learning techniques may be applied to the present invention depending on the embodiments.

[0047] FIG. 2a is a diagram briefly illustrating an indoor IoT system (200) according to one embodiment of the present invention.

[0048] An IoT system (200) according to one embodiment of the present invention may include one or more sensors (210-1 to 210-4) and a receiver (220) configured to be communicated with one or more sensors.

[0049] One or more sensors (210-1 to 210-4) according to one embodiment of the present invention may include a sensor unit for sensing one or more data, a processor for preprocessing the sensed data and converting the data, a communication unit configured to transmit the data processed by the processor to a receiving unit and receive a set value from the transmitting unit, and a memory for storing the sensed data and the processed data.

[0050] The sensor portion of one or more sensors (210-1 to 210-4) may include an indoor air quality sensor, a motion sensor, a temperature sensor, a pH sensor, a humidity sensor, a rotation sensor, a light intensity sensor, etc. The indoor air quality sensor may be configured to collect atmospheric environment information (PM10, PM2.5, CO, TVOC, HCHO, etc.). Meanwhile, the indoor air quality sensor may be docked and connected to the IoT system (200).

[0051] The processor of one or more sensors (210-1 to 210-4) may be configured to convert measured data based on a pre-stored setting value, etc. Such setting value may be converted by an external command to change the setting value.

[0052] The communication unit of one or more sensors (210-1 to 210-4) may be configured to transmit sensed data or data processed by a processor to a receiver and to receive a set value from a transmitter. Additionally, the communication unit may optionally communicate by wireless communication with networks, such as the Internet, also referred to as the World Wide Web (WWW), an intranet, and / or wireless networks, such as a cellular telephone network, a wireless LAN (local area network) and / or a MAN (metropolitan area network), and other devices. The communication unit of one or more sensors (210-1 to 210-4) may optionally include a well-known circuit for detecting near field communication (NFC) fields, for example by a short-range communication radio. Here, wireless communication optionally includes GSM (Global System for Mobile Communications), EDGE (Enhanced Data GSM Environment), HSDPA (high-speed downlink packet access), HSUPA (high-speed uplink packet access), EV-DO (Evolution, Data-Only), HSPA, HSPA+, DC-HSPDA (Dual-Cell HSPA), LTE (long term evolution), NFC (near field communication), Wideband code division multiple access (W-CDMA), code division multiple access (CDMA), time division multiple access (TDMA), Bluetooth, Bluetooth Low Energy (BTLE), Wireless Fidelity (Wi-Fi) (e.g., IEEE 802.11a, IEEE 802.11b, IEEE 802.11g, IEEE 802.Any of a plurality of communication standards, protocols, and technologies may be used, including but not limited to any other suitable communication protocols, such as 11n and / or IEEE 802.11ac), VoIP (voice over Internet Protocol), Wi-MAX, protocols for email (e.g., IMAP (Internet message access protocol) and / or POP (post office protocol)), instant messaging (e.g., XMPP (extensible messaging and presence protocol), SIMPLE (Session Initiation Protocol for Instant Messaging and Presence Leveraging Extensions), IMPS (Instant Messaging and Presence Service)), and / or SMS (Short Message Service), or communication protocols not yet developed at the time of filing of this document. The network interface of this communication unit is a subcomponent that enables data to be exchanged between external devices (e.g., receiver (220)) and other devices through one or more networks. For example, a network connection with an indoor router can be established via WIFI.

[0053] One or more sensors (210-1 to 210-4) may also include a display unit. Visual warnings can be displayed according to energy consumption through the display unit.

[0054] A receiver (220) according to one embodiment of the present invention may include a receiver (222) configured to receive IoT collected data from one or more sensors, a processor (224), and a non-transient memory (226) in which one or more programs are stored.

[0055] The receiver (220) may include a processor as a computing device and may be a device such as a computer, tablet computer, multimedia device, mobile device, server, network server, data server, etc.

[0056] The receiver (220) can typically support various applications such as data modification / processing applications, drawing applications, presentation applications, word processing applications, website creation applications, disk authoring applications, spreadsheet applications, game applications, telephone applications, video conferencing applications, email applications, instant messaging applications, exercise support applications, photo management applications, digital camera applications, digital video camera applications, web browsing applications, etc.

[0057] One or more programs may include instructions that, when executed by a processor, cause the processor to perform a method for generating data for artificial intelligence learning based on the name service.

[0058] A processor may be embodied using any suitable instruction set architecture (ISA) and configured to execute instructions defined within that instruction set architecture. The processor may be any of the various instruction set architectures (ISAs), such as x86, PowerPC, SPARC, RISC, ARM, or MIPS ISAs, or general-purpose or embedded processors using any other suitable ISA.

[0059] The processor may include neural processor circuits for artificial intelligence, which are circuits that perform various machine learning operations based on operations including multiplication, addition, and accumulation. These operations may be arranged, for example, to perform convolution between input data and kernel data. The neural processor circuits may be configurable circuits that perform these operations in a high-speed and power-efficient manner while alleviating resource-intensive operations associated with neural network operations from the central processing unit.

[0060] FIG. 2b is a drawing showing an example of collecting outdoor weather information according to one embodiment of the present invention.

[0061] According to one embodiment of the present invention, the UM-LDAPS (Unified Model-Local Data Assimilation and Prediction System), a numerical weather forecasting system operated by the Korea Meteorological Administration in FIG. 2b, may be used to collect outdoor weather information. FIG. 2b is a model for forecasting local weather phenomena on the Korean Peninsula, with a horizontal spatial resolution of 1.5 km and a vertical grid consisting of 70 layers from the ground to an altitude of approximately 40 km. That is, it is a model that calculates the average east-west, north-south, and vertical velocities, temperature, and water vapor content for a 1.5 km × 1.5 km space using the Navier-Stokes equations. Since the boundary conditions for the ground include not only topography but also soil temperature and moisture content based on soil composition, it is a model that reflects detailed regional characteristics of a dense area.

[0062] The Korea Meteorological Administration can produce forecast times ranging from 36 to 48 hours based on four analysis times per day (03, 09, 15, and 21 KST). Meanwhile, since a larger time difference between the analysis time and the forecast time can cause errors, it is common practice to use data with a difference of 0 to 5 hours.

[0063] FIG. 3 is a diagram briefly illustrating a machine learning learning method (300) according to one embodiment of the present invention.

[0064] A machine learning method (300) according to one embodiment of the present invention may include a step of obtaining power consumption information (S310), a step of preprocessing data to be learned in an artificial intelligence learning model (S320), and a step of selecting a model for AI learning (S330).

[0065] In the step of acquiring power usage information (S310), the power usage information may be sensing data. Acquiring sensing data may include the process of sensing specific data. For example, Advanced Metering Infrastructure (AMI) power usage information may be acquired as power usage information. In the case of the telecommunications big data platform (https: / www.bigdata-telecom.kr), Korea Smart Grid provides AMI power usage information (anonymized) based on apartment complexes at one-hour intervals from May 1, 2017, to November 2021, and after acquiring such data, the process may proceed to the preprocessing step.

[0066] The step of preprocessing data to be trained in an artificial intelligence learning model (S320) may refer to a process for refining the data before inputting it into the learning model. In other words, to construct AI-based training / validation data, spatiotemporal alignment and dynamic elements are required depending on the characteristics of the data. That is, the preprocessing step may be a preliminary operation to separate, remove, and process unnecessary information to treat sensed raw data into a form suitable for the purpose and method of data analysis. For big data analysis and data mining, well-prepared data according to the requirements of each algorithm is essential. Therefore, there is a need to effectively process collected data to suit the purpose, and to this end, measurement errors are reduced and noise, distortion, and deviation are minimized. To resolve issues such as precision, accuracy, outliers, missing values, contradictions, inconsistencies, and duplication, methods such as the following preprocessing techniques are used.

[0067] Data preprocessing involves processing, analyzing, and fusing data. First, a system environment can be established through data design, and metadata can be defined. Next, the data can be processed. At this stage, basic statistics can be analyzed, outliers can be removed, and inconsistencies can be corrected. Additionally, the data can be structured, transformed, and extracted, and the data can be set as training and test datasets. Next, the data can be analyzed. At this stage, power consumption based on customer location (periods such as 24 hours, weekly, monthly, seasonal, and annually) can be analyzed. Furthermore, correlations between power consumption and atmospheric environment / weather information can be derived. Next, the data can be fused. At this stage, meta-information between weather information (atmospheric environment, UM local model) can be analyzed, and data can be interpolated according to spatiotemporal resolution. As described later in Figure 4, this may include preprocessing of indoor atmospheric environment information, preprocessing of outdoor weather information, and the addition of dynamic elements (solar information, addition of indexing).

[0068] The step of selecting a learning model (S330) may include a step of selecting a suitable artificial intelligence learning model among several artificial intelligence learning models. Artificial intelligence learning models may include unsupervised learning models, supervised learning models, and hybrid learning models. For example, unsupervised learning models may include deep neural networks (DNN), self-organizing maps (SOM), adaptive resonance theory (ART), deep autoencoders, deep belief networks (DBN), Boltzmann machines (BM), restricted Boltzmann machines (RBM), K-means algorithms, etc. For example, supervised learning models may include hidden Markov models (HMM), perceptrons, convolutional neural networks (CNN), support vector machines (SVM), Naive Bayes classifiers (NBC), etc. For example, hybrid learning models are models proposed to solve the problem where learning efficiency can vary depending on the starting point of learning, as the training data is generally just measured data and does not include errors contained in the data or the correlation between each data point, and unsupervised and supervised learning start learning at arbitrary points. AI modeling can be performed through model development and model enhancement. Model development can be carried out by selecting optimal variables (dynamic / static elements), training machine learning (ML) / deep learning (DL) models, and predicting AI power consumption / saving amounts. Model enhancement can optimize hyperparameters every 24 hours and perform prediction and verification using real-time input data.Here, hyperparameters are variables determined by the user and differ from general parameters.

[0069] FIG. 4 is a diagram specifically illustrating a data preprocessing step (S320) according to an embodiment of the present invention.

[0070] In step S410, indoor air environment information can be preprocessed. Information within a radius of 1 km from the observation station can be utilized, but significant errors may occur in other cases. To prevent this, a 1 km grid can be constructed using point information (fine dust, carbon monoxide, etc.), and statistical spatial interpolation can be performed based on customer location-based latitude and longitude to improve the accuracy of the observation information.

[0071] In step S420, outdoor weather information can be preprocessed. For numerical weather prediction models, converting Grib2 (.grb2) files into text can result in slow I / O data access speeds and large file sizes. Therefore, for faster computation, it is common practice to preprocess by converting to the scientific and technological format NetCDF (.nc) using a single grid based on customer location-based latitude and longitude.

[0072] In step S430, dynamic elements may be added. Adding dynamic elements may include adding solar information and indexation.

[0073] According to one embodiment of the present invention, in order to improve the performance of short-term forecasting based on customer location, seven types of solar information were added in addition to indoor atmospheric environment and outdoor weather information. That is, generally, the solar zenith angle, solar azimuth angle, solar hour angle, and all-weather / direct / scattered solar radiation on clear days can be theoretically calculated using the customer location (latitude, longitude), date and time, and local time (KST). In addition, since climatologically, monthly atmospheric turbidity may exhibit hourly discontinuities, the present invention resolves this through prior interpolation.

[0074] To enhance the effectiveness of meteorological information from the aforementioned numerical weather prediction model, a direct correlation with power consumption can be established through the Temperature-Humidity Index (THI) and Wind Chill (WCT). In particular, the Wind Chill is defined to measure how cold people feel in generally cold and windy weather. It is calculated based on the lookup table in Fig. 5, based on tests conducted on humans, as a predefined function of temperature and wind speed variables representing the body's sensation on exposed skin.

[0075] The temperature-humidity index (THI) and the perceived temperature-humidity index (WCT) according to one embodiment of the present invention are as follows.

[0076]

[0077]

[0078] Here, TA, HM, and WS represent temperature (°C), relative humidity (%), and wind speed (m / s), respectively.

[0079] User-customized customer location-based training / validation data can be generated through the results of steps S410 to S430 mentioned earlier.

[0080] The data for the preliminary investigation to build the training / validation data is as follows.

[0081] Data Features Variables Time Time Resolution Spatial Resolution Training / Validation Data Indoor Air Environment (2 types) Fine Dust (PM10, PM2.5) Carbon Monoxide (CO) 1 Hour Specific Point Training Data 1 year 3 months (2017.07.01.~2018.09.27.) Validation Data 48 hours (2018.09.23. 23:00~2018.09.30. 23:00) 168 hours (2018.09.28. 23:00~2018.09.30. 23:00) 336 hours (2018.09.16. 23:00~2018.09.30. 23:00) Outdoor Weather (15 types) Total Cloud Cover (CA_TOT) Relative Humidity (HM) Atmospheric Pressure (PA) Temperature (TA) Dew Point Temperature (TD) Wind Direction (WD) Wind Speed ​​(WS) Solar Radiation (SWR) 602 x 781 (1.5 km) Solar Information (7 types) Solar Zenith (SZA) Solar Azimuth (AZA) Solar Hour Angle (ET) Atmospheric Turbidity (TURB) Clear Day Total Solar Radiation (GHI_CLR) Clear Day Direct Solar Radiation (DNI_CLR) Clear Day Scattered Solar Radiation (DHI_CLR) None Indexation (2 types) Temperature-Humidity Index (THI) Feels Like Temperature (WCT) Hourly Information (5 types) Weekend (WEEKDAY_YN) Spring (SPRING_YN) Summer (SUMMER_YN) Autumn (FAIL_YN) Winter (WINTER_YN) None AMI Power Consumption (1 type) 1-Hour Power Consumption Customer Location-Based Latitude & Longitude

[0082] The inventors of the present invention developed a power consumption algorithm through automated / manual learning of artificial intelligence based on the data preprocessed in Fig. 4. Generally, manual learning (AI development) utilizing data faces difficulties due to a lack of data scientists with specialized knowledge in algorithms or statistics, and to solve this, automated learning modeling (AutoML) can be utilized. Automated learning modeling (AutoML) does not require the expertise of a developer and can be developed by business or corporate practitioners with basic data analysis skills. It can automatically build a model with precision equal to or greater than that of a general data scientist by rationally combining more than 100 types of algorithms. Since AutoML models automatically, there is no need to understand business requirements, and it can be implemented within approximately one week. Furthermore, while open source software is free, commercial software has the advantage of being adjustable to scale through PoC (Proof of Concept).

[0083] Therefore, the present invention aims to provide a high-precision short-term prediction solution by performing an AI automated / manual learning modeling process.

[0084] FIG. 6 is a diagram showing the contribution according to input variables of manual learning modeling of short-term prediction according to one embodiment of the present invention.

[0085] FIG. 6 is a diagram (600) showing the contribution of the learning model during manual learning of 48-hour short-term prediction in daily rule patterns, with the contribution of the input variables arranged in descending order.

[0086] In one embodiment according to the present invention, manual modeling was constructed using LightGBM, a representative boosting-based machine learning technique, for manual learning (AI development) modeling.

[0087] LightGBM was introduced by Microsoft in 2017 as an algorithm designed to maintain performance while reducing modeling time by rapidly calculating information gain using only a portion of the data. It shortens data processing time, a drawback of other models utilizing existing boosting algorithms, thereby demonstrating good performance in terms of speed. LightGBM's tree is a vertically expanding leaf-wise structure, and by repeating the leaf node with the maximum loss value, it can minimize prediction error loss compared to balanced tree splitting methods.

[0088] In other words, a disadvantage of the GBRT (Gradient Boosted Regression Trees) algorithm is that computation time increases significantly as the amount of data increases, leading to a decrease in the efficiency of model training. To address this problem, we propose a method using GOSS (Gradient-based One-Side Sampling) to partially randomly sample samples with small gradients, while focusing entirely on learning for samples with large gradients. As a result, LightGBM is based on GOSS, which can reduce model complexity and perform training more efficiently.

[0089] In the manual learning process, the contribution according to the input data is specifically visualized in Figure 6, and it can be observed that the contribution is highest in the order of weather, atmospheric environment, and time information. Generally, since using input variables with low contribution increases the complexity of the learning model, the top 20 input variables (TURB~CO) in Figure 6 were used.

[0090] Furthermore, since manual learning performs prediction based on historical data using a single model (e.g., LightGBM), the short-term prediction algorithm contains inherent biases from the historical data. To address this, the present invention requires performing bias correction based on short-term prediction and actual measurement results.

[0091] According to one embodiment of the present invention, for bias correction, a short-term prediction result (PRD) based on the following simple linear regression equation and a bias-corrected prediction result (PRD) through arbitrary coefficients (a, b) are provided. biasCorr ) can be calculated.

[0092]

[0093] According to one embodiment of the present invention, a bias-corrected prediction result (PRD) based on actual power measurement (OBS) biasCorr The root mean square error (rMSE), which is a validation score for ), can be calculated. Therefore, if the bias correction process is repeated infinitely, as in the Monte Carlo method, the root mean square error according to arbitrary coefficients (a, b) can be simulated. Since the bias correction process cannot actually be repeated infinitely, a finite number of iterations can be set as n. For example, the number of iterations can be set to 100, 1000, 10000, etc. Finally, the a and b that minimize the root mean square error are found and defined as the bias correction coefficients.

[0094]

[0095] Meanwhile, a learning model for short-term prediction can be implemented using automatic learning (AutoML) modeling.

[0096] FIG. 7 is a diagram showing the contribution according to input variables of an automatic learning model of short-term prediction according to one embodiment of the present invention.

[0097] According to one embodiment of the present invention, automatic modeling can be constructed using FLAML, a representative automatic learning method. Such automatic learning (AutoML) modeling can be implemented in the following steps.

[0098] Step 1 (Model Type): You can select regression for data trends such as predicting power demand / savings.

[0099] Step 2 (Significant Variables): Independent variables (static variables: time information, dynamic variables: indoor air environment information, weather information) and dependent variables (electricity demand) can be selected as significant variables not only through their correlation with one another but also through variable selection techniques (forward selection, backward elimination, stepwise selection).

[0100] Step 3 (Data Splitting): Data can be split for training / validation, outliers removed, result values ​​supplemented and removed, categorical data converted into dummy variables, and numeric / categorical variables normalized.

[0101] Step 4 (Model Training / Optimization): You can select various machine learning / deep learning algorithm types (linear regression, random forest, XG Boost, etc.) and set the number of iterations and hyperparameters accordingly to perform training.

[0102] Step 5 (Model Validation): The most suitable model can be selected by evaluating the validation scores (correlation coefficient, root mean square error, etc.) for the selected algorithm.

[0103] Step 6 (Model Deployment): The optimal training model can be saved for real-time prediction services in web / app applications and open APIs.

[0104] Through this automatic learning, it was confirmed that the short-term forecast of power consumption exhibited not only low root mean square error and high correlation but also a trend similar to actual measurements during power peaks.

[0105] FIG. 8 is a diagram illustrating an artificial intelligence modeling learning method for power consumption prediction according to an embodiment of the present invention.

[0106] A method for learning artificial intelligence modeling for power consumption prediction according to one embodiment of the present invention can be performed by a computing device including a processor.

[0107] In step S810, indoor environment data of a specific household can be sensed.

[0108] In step S820, outdoor weather data for the area where a specific household is located can be received.

[0109] In step S830, power usage data of a specific household can be received.

[0110] In step S840, artificial intelligence modeling can be trained using indoor environment data, outdoor weather data, and power consumption data as parameters.

[0111] In step S850, the contribution of parameters to the power prediction accuracy can be calculated based on actual power measurements.

[0112] In step S860, the parameters can be determined using utilization parameters whose contribution is greater than or equal to a preset value.

[0113] In this specification, only a few examples among the various embodiments performed by the inventors are described; however, the technical concept of the present invention is not limited or restricted thereto, and it is obvious that it can be modified and implemented in various ways by those skilled in the art.

[0114]

[0115] According to the present invention, since it can be utilized as an energy management system for buildings as well as customer locations to achieve energy saving and rationalization of use, it has high industrial applicability.

Claims

1. A method for learning an artificial intelligence model of power consumption prediction performed by a computing device including a processor, A step of sensing indoor environment data of a specific household; A step of receiving outdoor weather data for the area where the aforementioned specific household is located; A step of receiving power usage data of the aforementioned specific household; A step of learning an artificial intelligence model using the indoor environment data, the outdoor weather data, and the power consumption data as parameters; A step of calculating the contribution of the above parameters to the power prediction accuracy based on actual power measurements; and Step of determining the utilization parameters among the above parameters whose contribution is greater than or equal to a preset value Power consumption prediction artificial intelligence modeling learning method including 2. In Paragraph 1, Bias correction is performed on the power prediction value (PRD) to obtain the bias-corrected prediction result value (PRD biasCorr Step to calculate ) A power consumption prediction artificial intelligence modeling learning method that further includes 3. In Paragraph 2, The actual power value (OBS) and the above bias-corrected predicted result value (PRD biasCorr A step of verifying bias-corrected prediction results by calculating the Root Mean Squared Error (rMSE) using ). Power consumption prediction artificial intelligence modeling learning method including further 4. In Paragraph 3, The above indoor environment data includes sensing values ​​for temperature, humidity, fine dust, ultrafine dust, carbon monoxide, carbon dioxide, total volatile organic compounds, and formaldehyde, respectively. The above weather data includes outdoor weather data, solar information data, indexed data, and time information data, in a method for learning an artificial intelligence model of power consumption prediction.

5. In Paragraph 4, The above indexed data includes the Temperature-Humidity Index (THI) and the Weather Condition Index (WCT), and A method for learning an artificial intelligence model for power consumption prediction, wherein the above temperature-humidity index (THI) and the above perceived index (WCT) are each calculated using the following formulas, and the following TA, HM, and WS are temperature (°C), relative humidity (%), and wind speed (m / s), respectively.

6. In Paragraph 4, The above outdoor weather data includes total cloud cover, relative humidity, atmospheric pressure, temperature, dew point temperature, wind direction, wind speed, and solar radiation, and The above solar information data includes solar zenith angle, solar azimuth angle, solar hour angle, atmospheric turbidity, total solar radiation on clear days, direct solar radiation on clear days, and scattered solar radiation on clear days, and The above indexed data includes the Temperature-Humidity Index (THI) and the Weather Condition Index (WCT), and The above time information data includes information regarding whether it is a weekend, whether it is spring, whether it is summer, whether it is autumn, and whether it is winter, a method for learning an artificial intelligence model of power consumption prediction.

7. In Paragraph 3, The above bias-corrected prediction result value (PRD biasCorr ) is calculated using the formula below, where a and b are arbitrary coefficients, and PRD is a short-term prediction result, an artificial intelligence modeling learning method for power consumption prediction.

8. In Paragraph 7, The above root mean square error is calculated using the following formula, and The OBS below is the actual power measurement value, and the PRD below biasCorr A power consumption prediction artificial intelligence modeling learning method, wherein is the above-mentioned bias-corrected prediction result value and n below is the number of bias correction processes.

9. In Paragraph 8, A power consumption prediction artificial intelligence modeling learning method in which the above-mentioned preset value is 20.

Images