Online prediction of weather-related overhead distribution line faults
The online failure prediction algorithm addresses data insufficiency and uncertainty in weather forecasts by dynamically updating predictive models using real-time data and Bernoulli Bandit problems, enhancing the accuracy and speed of power restoration in power distribution networks.
Patent Information
- Application Number
- JP2021183621
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2021-01-27
- Filing Date
- 2021-11-10
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-11-10
AI Technical Summary
Existing failure prediction methods for overhead power distribution lines face challenges due to insufficient data, imbalanced historical data, and inherent uncertainties in weather forecasts and individual variations, leading to inaccurate and non-updated predictive models, which result in delayed power restoration and increased customer dissatisfaction.
An online failure prediction algorithm that dynamically updates predictive models using real-time reported fault data, representing weather conditions in a tabular format and formulating the failure prediction problem as an online sequential decision-making problem with independent failure prediction models for each tabular cell, utilizing Bernoulli Bandit problems and machine learning frameworks to enhance prediction accuracy.
Enables faster and more cost-effective power restoration by accurately predicting component failures, reducing latency in data collection and model verification, and improving the resilience of power distribution networks against severe weather events.
Smart Images

Figure 0007584391000036 
Figure 0007584391000037 
Figure 0007584391000038
Abstract
Description
[Technical field]
[0001] Field The present invention relates to power systems, and more particularly to an online fault prediction algorithm for components including power distribution lines that utilizes real-time reported incident data to dynamically update predictive models. [Background technology]
[0002] background The electric grid is one of the most critical infrastructures for society today; for example, transportation, water supply, schools, government offices, and airports all depend on the supply of electricity. The electric power system infrastructure that provides the power includes electric lines, transformers, and other devices for generating, transmitting, and distributing electricity. Unfortunately, the increasing frequency, duration, and intensity of extreme weather events poses significant threats to the overhead distribution lines of the electric power grid, resulting in widespread blackouts primarily affecting the low-voltage and medium-voltage electric power grids, which contribute to the majority of the blackouts. For example, the United States in August and September 2017 was hit by three separate hurricanes, Harvey, Irma, and Maria, which resulted in the loss of power to a total of approximately 7.5 million customers in Texas and Florida. The enormous cost of power system incidents and their impact on personal safety have created a critical need to build resilience into the electric power system in light of today's aging and fatigued electric power grid. One effective way to enhance the operational resilience of the power grid is to achieve rapid and accurate estimation or prediction of faults occurring in the power grid. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] US2014 / 025,7913 A1 Summary of the Invention [Problem to be solved by the invention]
[0004] Some fault prediction methods for electric power distribution systems can include predicting the total number of power line faults, or the total duration of faults for a particular area within a particular time interval. However, these fault prediction methods face significant technical challenges when attempting to build a fault prediction model. For example, the first challenge may be not having the data available to make the prediction. The second challenge is that there needs to be data that can be used to derive individual fault models for each distribution line, as an example, to predict individual overhead distribution line faults. For example, using a method based on logistic regression based on events in the historical days. However, in most of the historical days, the fault statuses of these overhead distribution lines are labeled as "in service", and there is a significant imbalance in the operational status of overhead distribution lines.
[0005] Another challenge to overcome is that of model updating. Most accident prediction models are derived based on past knowledge or historical data. However, in practice, past knowledge is not enough evidence to support current decisions or predictions. For example, at least one reason is that most weather conditions, especially some extreme weather events, may be difficult to predict and / or rare in occurrence, lacking sufficient accident data to be used in accident prediction models. Also, the distribution of accident data is not necessarily uniform. The inherent uncertainty of weather forecasts and individual differences (e.g., vegetation environment, pruning, aging, etc.) can change the integrity of any historical accident data.
[0006] Traditionally, outage analysis relies on customer telephone calls made to utility companies in the event of an outage as the primary source of information for such analysis. This process can be very slow because many customers may not call to report an outage, and those who do report an outage may wait a relatively long period of time to report the outage, often assuming that a neighboring customer will call instead.
[0007] When a customer contacts a utility to report an outage, a repair crew is dispatched to determine the faulty grid element, perform the fault repair, and ultimately restore service. Without the aid of details about the extent of the damage and the location of the damage causing the outage, the repair crew may not be sure if the outage is a single location outage or a multi-location outage. However, repairs are not performed until the customer experiences a power outage and takes steps to contact the utility. And, as described above, the customer is left without power until a repair crew is dispatched, the cause of the outage is determined, and power is restored, which leads to poor service quality and increased customer dissatisfaction.
[0008] There are several accident or failure prediction methods for overhead distribution lines for power systems. For example, the method of US2014 / 025,7913 A1 predicts the accident status of overhead distribution lines using a failure rate model. The failure rate model uses historical data to model the number of times an event occurs in a time or space interval by using a statistical approach (such as a generalized linear regression model) or a machine learning technique (such as a support vector machine approach) based on historical accident samples. However, the method of US2014 / 025,7913 A1 has many problems including a lack of data available to make accurate predictions. To predict individual line accidents, a method needs to derive individual failure models for each distribution line, which can be by using a method based on logistic regression based on historical date events. However, most accident status reports for distribution lines by the power grid system are labeled as "in service", and there is a significant imbalance in reporting actual distribution line service status with the data used in the method of US2014 / 025,7913 A1. Also, incident data for overhead distribution lines is not necessarily uniform, as in the US 2014 / 025,7913 A1 method. For example, there are inherent uncertainties in weather forecasts and individual distribution line variations (e.g., vegetation conditions, age, etc.) that can alter the reliance on historical incident data, as well as the data used in the US 2014 / 025,7913 A1 method.
[0009] Although existing fault prediction methods may solve some specific incident prediction problems, they do not solve the online fault prediction problem. Therefore, there is a need for an online fault prediction method for overhead distribution lines for power distribution system restoration that includes dynamic updating of a predictive model utilizing real-time reported incident data.
[0010] overview The present invention relates generally to power systems, and more particularly to an online fault prediction algorithm for components including power distribution lines that utilizes real-time reported incident data to dynamically update predictive models. [Means for solving the problem]
[0011] Embodiments of the present disclosure include an assessment approach to analyze the impact of weather events on overhead distribution line or conductor fault status in an electric power distribution system. Some embodiments include online fault status prediction for overhead distribution lines, which are the most vulnerable components in an electric power distribution system. Historical data of past weather events, such as hurricanes or other severe events, i.e., mixed snow / rain events, tropical cyclones with winds of 39-73 mph, thunderstorms, snow events, etc., may be collected and used in fault prediction modeling.
[0012] At least one realization of an embodiment of the present disclosure includes representing weather conditions in a tabular format to aid in formulating a fault prediction problem. This can help identify a relationship between two weather variables, i.e., maximum instantaneous wind speed and lightning current, or any other two or more variables that convert each weather variable domain (minimum and maximum values) into a tabular format. Essentially, the minimum and maximum values are divided into a range of a set of cells in the tabular format, with each cell in the set representing a unique climatic variation of the relationship between the two weather variables. Each cell is then compared to a threshold value of a particular amount of maximum instantaneous wind speed and an amount of lightning current to identify a grouping of cells that meets the threshold value that accurately represents the weather condition scenario in a particular area within the region where the overhead distribution line may be located. If the region is divided into a grid arrangement of cells, and a cell that meets the threshold value is adjacent to other similar cells, when displayed, the particular area containing the group of cells accurately represents the weather condition scenario. Some advantages of representing weather variables in a tabular format may include reducing the amount of weather condition scenarios from entire areas to specific groupings of cells, which allows for no need to incorporate forecast accuracy techniques compared to traditional methods that use individual values (min / max) associated with entire areas.
[0013] Another realization is that using historical accident data in forecasting models alone is not enough evidence to support grid operators' real-time or current-time decisions in predicting distribution line failures. This is because the historical data is not directed to extreme weather events, lacks accident data to predict component failures, and is unreliable. At least one solution to this problem may be to reformulate the failure prediction problem as an online sequential decision-making problem, where an independent failure prediction model is used for each tabular cell, i.e., each failure predictor in the problem attempts to solve multiple independent Bernoulli bandit problems.
[0014] An embodiment of the present disclosure includes a failure predictor that accurately predicts the status of a component upon receiving forecasted or observed weather conditions. Observed data is data received via a communication network or wired connection and may include weather data from a local network of weather stations, third party weather forecasts, or any other weather related entity. Observed data includes component data from sensors associated with the component that communicate with a power grid system or an electric utility company's supervisory control and data acquisition (SCADA) system, which monitors and relays information associated with the component, i.e., power line status and other infrastructure status. For example, a remote terminal unit (RTU) may connect to a sensor, convert the sensor signal to digital data, and transmit the digital data to a system of the SCADA system, which is connected to a communication network, which communicates to the component failure prediction system. Most control operations of the SCADA system may be performed by an RTU or by a programmable logic controller (PLC). At the RTU or PLC level, the PLC may control some processes so that data may be collected from other systems, including equipment status reports, which are communicated to the SCADA system via the communication network.
[0015] Prediction can be achieved by predicting the status of a component as in service or accident, with each action associated with a probability of prediction accuracy that can be modeled by a Bernoulli distribution. The Bernoulli bandit problem involves choosing an action among multiple independent competing actions and selecting the action that receives a reward, with the reward associated with a probability in the Bernoulli distribution. The objective of the multi-armed bandit problem is to use the best strategy to select the action that maximizes the expected reward function. By utilizing the strategies provided by the multi-armed bandet problem, failure prediction can be solved to achieve the highest long-term accuracy for predicting component failures.
[0016] For example, first, the parameters of the failure prediction model for each weather range combination can be determined using the output of the forecast model of the historical weather events that fall into the corresponding range and updated with the output of the forecast model of the impending weather. Failure prediction can be treated as an online learning process of a multi-round repeated game. In each round, a prediction is first made, then observations are obtained to reveal the true outcome, and finally, the corresponding reward or loss is determined based on the accuracy of the prediction. The objective is to maximize the reward or minimize the loss, which is defined based on the cumulative reward / loss and the cumulative reward / loss of the best prediction strategy. Some of the reasons for this are that during experiments, it was found that the accident status of each component can be predicted directly using online sequential decision making. Meanwhile, by updating the prediction model online, the effects of existing prior conditions (age, wear conditions, etc.) can be incorporated and real-time changes in the surrounding conditions (falling trees, etc.) can be easily reflected.
[0017] The next step of some embodiments deals with minimizing the impact of unknown inherent uncertainties on the failure prediction model accuracy. These uncertainties may arise from weather forecasts and individual differences (e.g., vegetation environment, pruning, aging, etc.). Weather forecast uncertainties are partially mitigated by presenting weather conditions by ranges rather than values. Uncertainties from individual differences are handled by selecting an appropriate online learning framework for failure prediction. Through the learning process, frameworks with stronger learning capabilities can quickly pick up the impact of unknown inherent uncertainties by updating the prediction model with the reported accuracy for the latest prediction.
[0018] For example, two types of learning frameworks can be incorporated into the design of a desired failure predictor, one based on maximum likelihood estimation and the other based on maximum a posteriori estimation. The maximum likelihood based learning framework estimates the parameters of the Bernoulli probability distribution, i.e., the mean of the distribution, by maximizing an action value function defined as the expected reward of failure prediction, so that the observed data is most probable under the assumed statistical model. The maximum likelihood based learning framework is a greedy method, which may converge to a suboptimal solution. To overcome the suboptimality, the ε-green algorithm can be used by randomly selecting an action between two actions with probability ε and selecting the action that maximizes the expected reward with probability (1-ε). Another variation of maximum likelihood estimation that can be used for failure prediction is the upper confidence bound (UCB) algorithm, which always selects the most greedy action to maximize a confidence upper bound defined as the expected reward plus a term that favors actions that have a strong possibility of having an optimal value. The maximum a posteriori based learning framework for failure prediction estimates the parameters of the Bernoulli probability distribution under the condition that a prior distribution using a beta distribution is given.
[0019] Two types of learning frameworks are designed to adjust the online model parameters with received reported accident data to mitigate the effect of unknown inherent uncertainties on the failure prediction model accuracy. The learning frameworks realize two tasks: one is to perform a component status prediction based on a probability distribution model of the prediction correctness, and the other is to update the parameters of the probability distribution model with the latest predictions and real-time reported statuses of the components. The learning framework is selected by a trade-off between computational load and estimation accuracy. The parameters can be stored in a database; thus, model updates can be performed by evaluating the stored parameters and updating them only with the latest predictions and corresponding status reports.
[0020] Another embodiment of the present disclosure includes a component failure prediction system for components including overhead distribution lines located in a predetermined area. Some examples of overhead distribution line or power line faults may include broken lines (i.e., wires), downed / broken utility poles, shorted online segments, malfunctioning line switches, and shorts on power feed transformers. The predetermined area is an area served by the power distribution system and includes a collection of components associated with the overhead distribution lines. The hardware or electronic processor can be used to access data from a data storage device, including historical data including weather and component data such as given weather variables (also referred to as a "set of weather drivers") associated with maximum instantaneous wind speeds and lightning currents corresponding to model forcings, and weather events corresponding to component incidents for the predetermined area. The weather can include data including one or a combination of seasonal data; classification of impending weather events as one of snow, ice, sleet, rain / flood, wind, windstorm / lightning, ambient air temperature, wildfire, hurricane, severe weather rating, severe windstorm rating, other mixed events, and the like. The components may include data for each component including one or a combination of: age; material type; specified type and design; mounting and support equipment; electrical equipment; vegetation assessment; pruning assessment; foliage assessment; installation and maintenance log; severe weather exposure log; assessment of vegetation and tree growth around overhead power lines. Here, by using historical weather data and historical component data including weather events corresponding to component incidents for a given area, a model forcing intensity group may be defined using variables of the historical weather events, and a machine learning (ML) model may be configured for each model forcing intensity group. Further, a set of parameters specific to the impending weather event information for a predefined location may be generated, which may include location, time of year, etc., and the parameters may include seasonal data for the season; classification of snow, ice, rain, sleet, wind, lightning, ambient air temperature, vegetation, land use, topography, etc. The parameters of the IW events may be used for resource allocation, component repair, and service restoration. The hardware processor of the disclosed embodiment may implement several steps including: ● Defining model forcing intensity groups using historical meteorological event variables and constructing a machine learning (ML) model for each model forcing intensity group; ● searching for weather events through the historical data, identifying a model forcing intensity group for the weather events, generating estimated output values using an ML model corresponding to the group, updating the ML model with the estimated output values and the historical component data, and repeating sequentially for each weather event; ● receiving continuous real-time observational data from an imminent weather (IW) event communication network over a plurality of time periods, and iteratively, for each time period; • Identifying model forcing intensity groups for IW events via observed meteorological data; ● Updating the ML models for that group with the latest predicted output values and observed component data; ● For the updated ML model, generating updated output values for each component over the time period that predict component failure or no component failure.
[0021] Aspects of the present disclosure include receiving real-time reported accident data via a communication network controlled by a centralized distribution control system of the power grid or via local communication devices located around the component under review. The effectiveness of algorithms with different action selection strategies can be evaluated using a set of weather event samples generated by simulation or collected from historical data, and the prediction accuracy can be measured using an accuracy metric for all samples and an F1 score metric for distinguishing between positive and negative samples, i.e., a binary accident status. The failure prediction model can be first trained using historical data and then used for real-time application, while continuing to update with the latest predictions and reported data.
[0022] Practical Applications The present embodiment solves the technical problem of predetermining the number of repair crews required locally or non-locally along with arranging crew reserves and outage equipment. By being better prepared, utilities can restore power faster and more cost-effectively, and instill confidence in users by knowing that the utility will make storm preparation decisions on actual quantitative analysis rather than making assumptions using historical data. The features of the presently disclosed embodiments arise from reformulating the failure prediction problem as an online sequential decision problem using historical data, and even real-time observational data, with independent failure prediction models for each tabular cell, i.e., each failure predictor attempts to solve multiple independent Bernoulli bandit problems in the problem.
[0023] For example, an embodiment of the present disclosure identifies the relationship between two weather variables (maximum instantaneous wind speed and lightning current) from traditional historical and real-time formatted data and converts each weather variable range (minimum and maximum values) into a table format. Essentially, the minimum and maximum values are divided into a range of a set of cells in the table format, with each cell in the set representing a unique climate variation of the relationship between the two weather variables. Each cell is then compared to a threshold of a specific amount of maximum instantaneous wind speed and a threshold of a certain amount of lightning current to identify a grouping of cells that accurately represents a weather condition scenario. Representing weather variables in a table format contributes to the technical effect of reducing the amount of weather condition scenarios from a whole range to a specific grouping of cells, which can meet technical objectives including no need to incorporate forecast accuracy techniques compared to traditional methods that use individual values (minimum / maximum values) associated with a whole range. More importantly, an embodiment of the present disclosure provides more accurate forecasts than traditional methods based in part on dividing traditional historical and real-time data from value formatted data into a range format into a grid-like arrangement of cells. The cells that meet the threshold are then adjacent to other similar cells, and when displayed, the particular area containing the group of cells accurately represents the weather condition scenario. The embodiments of the present disclosure solve the technical problem of predetermining the number of repair crews required locally or non-locally, along with arranging crew reserves and outage equipment, using the resulting output values to predict the status of the components for that period of time, whether they are faulty or not. Allowing utilities to be better prepared, resulting in faster power restoration and increased cost-effectiveness. The reason this technical problem needs to be addressed is because severe weather causes significant damage to the distribution network and causes outages. Managing the risk of hazardous weather for utility engineers and emergency preparedness staff requires receiving accurate information using time series weather forecasts for predicted incidents, which the embodiments of the present disclosure provide.
[0024] Experimental Method Experiments have shown that representing weather conditions in tabular form can provide multiple advantages with respect to failure prediction. The first advantage is that it simplifies the complexity of the failure prediction model by restricting the weather condition variations to a set of sections in a domain that correspond to the inherent climatic variations of the weather within a range. The second advantage is that it ensures that there is sufficient data to determine the failure prediction model.
[0025] In developing the failure prediction model of the present disclosure, some realizations from the experiments are that expressing weather conditions in a tabular form is not important and is a unique feature. This is because the complexity of the relationship between component failures and weather conditions is further defined or narrowed to a set of sections in a domain within a specific range, i.e., within the inherent climate variability of weather within a range. By significantly reducing the complexity of the relationship between component failures and weather conditions to a set of sections in a domain compared to the conventional method having a much broader range, a more simplified model can be used for failure prediction to meet the computation speed requirements of online applications that can meet today's forecasting demands within the weather forecasting industry.
[0026] Continuing with the experiment, the impact of weather conditions is evaluated by a range when using the tabular representation instead of the traditional approach of using individual weather condition values. For example, traditional power line incident prediction methods analyze the impact of weather events on a wider range when weather conditions are formatted / represented as individual values. The conversion of weather conditions to a range by the tabular representation of the present embodiment is a more simplified model that can relate component failures to weather conditions. Some reasons why the tabular representation is more simplified are that the weather condition difference for weather events that belong to the same range defined by the tabular representation can be ignored.
[0027] Some aspects of why the value-to-range conversion is so different from the traditional approach of using values, as opposed to using values by traditional forecasting methods, may lie in the configuration of how the forecasting problem is solved. As discussed above, without accurate historical data, it is difficult for a fault prediction model to predict overhead distribution line faults. One surprising benefit later realized from representing weather conditions in a table is the possibility of reducing the amount of weather condition scenarios and even the requirement of forecast accuracy, as mentioned above. The reason why it is important to reduce the amount of weather condition scenarios in a fault prediction model is that when weather conditions are represented in a tabular format, detailed values are not required for each weather variable as in the past, since the difference between weather events belonging to the same tabular format cell is ignored. By doing so, the power utility can reduce the wait time for more accurate and higher resolution weather observations and can start preparations and command duties for disaster repair and recovery much sooner when component failure information is available. Another advantage over using a tabular representation is that by representing weather conditions as ranges instead of values, it becomes much easier to collect a sufficient number of past weather events to configure and calibrate a component failure model. Therefore, the difficulty of data creation and model validation for online failure prediction is also reduced.
[0028] Further testing included formulating the failure prediction problem as an online sequential decision problem with an independent failure prediction model for each tabular cell, i.e., combination of weather variable ranges. It was found that the online sequential decision approach of the present disclosure is very different from some experimentally tested failure prediction approaches. For example, a first test approach included predicting the total number of track incidents and the total duration of incidents for a specific area in a specific time period using historical weather data. However, this first test approach merely provided a statistical value prediction of the total incidents and total duration, but did not indicate the location and time of the incident occurrence, which is important for the power distribution system to perform fault isolation, restoration, and component repair operations. Finally, after experimentation with the first test approach, testing was stopped due to failure predictions that could not be obtained that met the minimum goals of the present disclosure. For example, the first test approach could not identify which component was experiencing an incident and when the incident was occurring. Following the first test, a second test approach was developed to be more fully vetted, which included incorporating a different statistical approach to achieve more successful failure predictions.
[0029] The second experimental approach involved the use of a failure rate model, also called a cluded fragility curve, used to provide a damage assessment for power equipment from weather events. This model assumed the number of accidents using historical accident sample data, and then used a Poisson distribution to model the number of times an accident event occurred within a time or space interval. To fully explore all possible solutions to develop a successful failure prediction approach, several different statistical approaches were incorporated along with the failure rate model, including machine learning techniques such as generalized linear regression models, support vector machine approaches based on historical accident samples.
[0030] However, what was later learned from the testing of the first and second test approaches was that building a fault prediction model faces significant technical challenges. The first challenge was that there was no data available for developing a fault prediction model to meet the goals of the present disclosure. For example, when predicting individual line faults, an individual fault model needs to be derived for each distribution line, for example, one approach may include using a method based on logistic regression based on historical day events. However, today's recording practice of reporting fault history by power system companies and third parties records the status of a distribution line during an incident as having an incident state labeled as "in service", and there is a serious imbalance in recording the operational status of a distribution line. The second challenge learned from the experiment was that the first and second test approaches as well as multiple other test approaches tested failed to include any aspect of updating the fault prediction model. These tested test approaches, and even conventional fault prediction approaches or conventional fault prediction models, were all derived based on past knowledge or historical accident samples.
[0031] What was realized from this experiment is that using only past knowledge, i.e., past accident data, is not enough evidence to support real-time or current in-time decisions in predicting the failure of the power grid or distribution lines or other components of today's power industry. First, historical recorded weather conditions are rarely directed towards extreme weather events, and historical recorded weather lacks sufficient accident data with respect to predicting power line failures of other components of the power grid system. Second, the distribution of accident data is not always uniform or reliable. The inherent uncertainty of weather forecast differences and individual differences (e.g., vegetation environment, aging, etc.) changes the confidence in the historical accident data. Therefore, by formulating the failure prediction problem as an online sequential decision problem, the failure prediction can not only learn from past knowledge and conditions, but also update correspondingly with the influence of real-time weather data.
[0032] In formulating the failure prediction problem as an online sequential decision problem, one set of test experiments includes reformulating the failure prediction problem, where each failure predictor in the failure prediction problem attempts to solve multiple independent Bernoulli bandit problems. A desired failure predictor accurately predicts the status of a component upon receiving forecasted or observed weather conditions. Prediction can be achieved by choosing to either predict the component status as in-service or as an accident, where each action is associated with a probability of prediction accuracy that can be modeled by a Bernoulli distribution. In this way, failure prediction can be viewed as a Bernoulli two-armed bandit problem. A Bernoulli multi-armed bandit problem is to select an action among multiple independent competing actions, where a reward is received upon selecting an action, where the reward is associated with a probability with a Bernoulli distribution. The objective of the multi-armed bandit problem is to use the best strategy to select the action that maximizes an expected reward function. By utilizing the strategies provided by the multi-armed bandit problem, failure prediction can be solved to achieve the highest long-term accuracy for prediction.
[0033] According to another embodiment of the present disclosure, a component prediction system for a region includes a component prediction processor configured to receive data including weather and component data for a region from a data store. Generate a dataset for the region to identify weather events with component locations, each including a parameter, and a set of weather variables converted into a tabular format. Generate a set of weather drivers or a set of model weather drivers, each set of weather drivers including a given set of weather variables. Use the dataset to construct a machine learning (ML) model for each set of weather drivers for each component in the region. For each weather event in the region, iteratively identify for each component a set of corresponding weather drivers with weather variables including maximum instantaneous wind speed and lightning current. Generate output values for the ML model corresponding to the identified components for the set of weather variables. Update the ML model with the output values and the component data for the weather events. Receive real-time observational data over a period of time via a communication network of impending weather (IW) events in the region. Iteratively identify for each period of time for each component a set of corresponding weather drivers with weather variables including maximum instantaneous wind speed and lightning current. The ML model corresponding to the identified component for the set of weather variables is updated using a last run iteration forecast output value for the component and the observational data of the IW event. For the updated ML model, an output value is generated that predicts a component status as failed or not failed for the time period.
[0034] According to another embodiment of the present disclosure, a component forecasting method for components in a region includes accessing data including weather and component data for the region. Generate a dataset for the region to identify weather events with component locations, each including a parameter, and a set of weather variables converted into a table format representing a grid of weather cells for the region. Generate a set of weather drivers, each including a given set of weather variables. Use the dataset to construct a machine learning (ML) model for each set of weather drivers for each component in the region, and iteratively identify, for each weather event in the region, a corresponding set of weather drivers for each component, along with weather variables including maximum instantaneous wind speed and lightning current. Generate output values for the ML model corresponding to the identified components for the set of weather variables, and update the ML model with the output values and the component data for the weather event. Receive real-time observational data for a period of an impending weather (IW) event in the region, and iteratively identify, for each period, a corresponding set of weather drivers for each component, along with weather variables including maximum instantaneous wind speed and lightning current. The ML model corresponding to the identified component for the set of weather variables is updated using a last run iteration forecast output value for the component and the observational data of the IW event. For the updated ML model, an output value is generated that predicts a component status as failed or not failed for the time period.
[0035] According to another embodiment of the present disclosure, a method for a computer program product for predicting overhead power line status of an electric power grid system in an area includes one or more computer readable storage media. The program instructions stored on at least one of the one or more computer-readable storage media include: (a) receiving data from a data storage device including weather and component data for a region, and generating a dataset for the region to identify weather events each including a parameter along with a component location, and a set of weather variables converted into a tabular format representing a grid of weather cells for the region, where for each grid, each weather cell represents a unique climate variability of relationships between the weather variables of the set of weather variables; (b) generating a set of weather drivers, each including a given set of weather variables, and constructing a machine learning (ML) model for each set of weather drivers for each component in the region using the dataset, and iteratively, for each weather event in the region, (c) identifying for each component a corresponding set of weather drivers along with weather variables including maximum instantaneous wind speed and lightning current; and (d) generating output values for the ML model corresponding to the identified components for the set of weather variables, and updating the ML model with the output values and the component data for the weather events. The program instructions stored on at least one of the one or more computer-readable storage media (a) receive real-time observational data over a time period via a communication network of an impending weather (IW) event in the region, and iteratively, for each time period, (b) identify from the observational data a corresponding component and a model forcing group of the IW event, (c) for each component, identify a corresponding set of weather drivers along with weather variables including maximum instantaneous wind speed and lightning current, and update the ML model corresponding to the identified component for the set of weather variables using a last performed iteration forecast output value for the component and the observational data of the IW event, and (d) generate an output value for the updated ML model that predicts a component status as failed or not failed for the time period.
[0036] According to another embodiment of the present disclosure, a method for a forecasting system for components in a region includes a forecasting processor configured to receive weather and component data for the region from a memory, generate a dataset for the region to identify weather event parameters along with component locations, generate a set of weather drivers, each set of weather drivers including a given set of weather variables converted into a tabular format representing a grid of weather cells configured to accurately represent weather condition scenarios in the region. Using the dataset, construct a machine learning (ML) model for each set of weather drivers for each component, and iteratively, for each weather event, (a) identify, for each component, a set of corresponding weather drivers along with weather variables, generate output values for the ML model corresponding to the identified component for the set of weather variables, and update the ML model with the output values and the component data for the weather event; (b) receive real-time observation data over a time period via a communication network of impending weather (IW) events in the region, and iteratively, for each time period, (c) identify, for each component, a set of corresponding weather drivers along with weather variables, and update the ML model corresponding to the identified component for the set of weather variables with last performed iteration forecast output values for the component and the observation data of the IW event, and generate an output value for the updated ML model that predicts a component status as failed or not failed for the time period.
[0037] Embodiments of the present disclosure will now be further described with reference to the accompanying drawings, in which: The drawings shown are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of the embodiments disclosed herein. [Brief description of the drawings]
[0038] [Figure 1A] FIG. 1 is a block diagram illustrating a component failure prediction method for a component, such as an overhead line, in an electrical distribution system located in a given area, according to an embodiment of the present disclosure. [Figure 1B] FIG. 1B is a block diagram illustrating a component failure prediction system, similar to the method of FIG. 1A, configured with several components, in accordance with an embodiment of the present disclosure. [Figure 1C] 1B is a schematic diagram illustrating a component failure prediction system, such as the method of FIG. 1A, according to some embodiments of the present disclosure, showing a central computing application connected to computer applications and computational processing and data storage. [Figure 2A] FIG. 1 is a schematic diagram illustrating components of an electrical power distribution system and the relationships between the system components, according to some embodiments of the present disclosure. [Figure 2B] FIG. 1 is a block diagram illustrating several method steps for an online power distribution line fault prediction problem, according to an embodiment of the present disclosure. [Figure 2C] FIG. 2 is a schematic diagram illustrating a telephone company landscape 200C in which some embodiments are configured to predict telephone component status susceptible to severe weather events, according to an embodiment of the present disclosure. [Figure 2D] FIG. 1 is a schematic diagram illustrating a construction project for which some embodiments are configured to predict construction project component status susceptible to severe weather events, in accordance with embodiments of the present disclosure. [Figure 3A] FIG. 2 is a schematic diagram illustrating a tabular weather representation according to an embodiment of the present disclosure. [Figure 3B] FIG. 1 is a schematic diagram illustrating deriving weather conditions for a smaller area based on larger area measurements according to an embodiment of the present disclosure. [Figure 4] FIG. 2 is a block diagram illustrating some method steps of a transformed online power distribution line fault prediction problem according to an embodiment of the present disclosure. [Figure 5A] FIG. 1 is a schematic diagram illustrating a Bayesian network for data generation according to an embodiment of the present disclosure. [Figure 5B] FIG. 2 is a block diagram illustrating several process steps for generating a dataset of artificial data points for fault prediction algorithm testing according to an embodiment of the present disclosure. [Figure 6A] FIG. 1 is a graph illustrating aspects of a failure rate model, according to an embodiment of the present disclosure. [Figure 6B] FIG. 1 is a graph illustrating aspects of a failure probability model, according to an embodiment of the present disclosure. [Figure 7A] 1 is a table illustrating parameter settings for wind-related data generation according to an embodiment of the present disclosure. [Figure 7B] 1 is a table illustrating parameter settings for lightning-related data generation according to an embodiment of the present disclosure. [Figure 8] FIG. 11 is a graph showing data distribution of artificial data points for maximum instantaneous wind speed and lightning current, a data set having 10,000 artificial data points is generated to validate the online forecasting algorithms described in Algorithms 1-2 and their variations according to an embodiment of the present disclosure. [Figure 9A] FIG. 1 is a graph showing evaluation results for five different algorithms using average precision, according to an embodiment of the present disclosure. [Figure 9B] FIG. 1 is a graph showing evaluation results for five different algorithms using average F1-score, according to an embodiment of the present disclosure. [Figure 10A] FIG. 13 is a graph illustrating the density of a probability distribution for the probability of reward 1 for giving a status prediction of 0 for a high occurrence weather event, according to an embodiment of the present disclosure. [Figure 10B] FIG. 13 is a graph illustrating the density of a probability distribution for the probability of a reward of 1 for giving a status prediction of 0 for a medium occurrence weather event, according to an embodiment of the present disclosure. [Figure 10C] FIG. 13 is a graph illustrating the density of a probability distribution for the probability of reward 1 for giving a status prediction of 0 for a low occurrence weather event, according to an embodiment of the present disclosure. [Figure 11A] FIG. 13 is a graph illustrating the density of a probability distribution for probability 1 of reward for giving a status prediction 1 for a high occurrence weather event, according to an embodiment of the present disclosure. [Figure 11B]FIG. 13 is a graph illustrating the density of a probability distribution for the probability 1 of reward for giving a status prediction 1 for a medium-occurrence weather event, according to an embodiment of the present disclosure. [Figure 11C] FIG. 13 is a graph illustrating the density of a probability distribution for the probability 1 of reward for giving a status prediction 1 for a low occurrence weather event, according to an embodiment of the present disclosure. [Figure 12] FIG. 1C is a block diagram illustrating the method of FIGS. 1A and 1B, which may be implemented using an alternative computer or processor, according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0039] Although the above-identified drawing figures set forth embodiments disclosed herein, other embodiments are contemplated, as described in the discussion. This disclosure presents exemplary embodiments by way of example and not limitation. Those skilled in the art can devise numerous other modifications and embodiments that fall within the scope and spirit of the principles of the disclosed embodiments.
[0040] Detailed Description The present disclosure relates to providing systems and methods for online fault prediction of components of an electrical power distribution system.
[0041] 1A is a block diagram illustrating a method 100A for predicting faults for components, such as overhead lines, in an electric power distribution system located in a given region, according to an embodiment of the present disclosure. Step 110 includes accessing data in step 115, the data including given weather variables related to maximum instantaneous wind speed and lightning current corresponding to model forcings (also referred to as a "set of weather drivers") and historical data including weather event data corresponding to component incident data for the given region. Continuing with step 110, the variables of the historical weather events are used to define model forcing intensity (FI) groups, and a machine learning (ML) model is configured for each FI group.
[0042] Step 120 of FIG. 1A includes sequentially retrieving the historical weather (HW) events from step 115 from the historical data access, identifying an FI group for the event, generating estimated output values using an ML model corresponding to the FI group, and updating the ML model with the estimated output values and the historical component data.
[0043] 1A includes receiving continuous real-time observational data from a communication network of an impending weather (IW) event for each time period over a plurality of time periods from step 130. Step 125 continues by identifying an FI group for the IW event from the observed weather data.
[0044] Step 135 of FIG. 1A involves updating the ML model using the last iteration predicted output values and the observed component data from step 130.
[0045] Step 145 of FIG. 1A includes generating updated output values for the updated ML model, for each component, for the time period, that predict component failure or no component failure.
[0046] 1B is a block diagram illustrating a component failure prediction system similar to the method of FIG. 1A, configured with several components, according to an embodiment of the present disclosure. One or more computers 102 communicate with an input interface 106, a memory 108, a communication network 104, one or more processors 116, and a device 123. The computer 102 is connected to a component 112, i.e., an overhead line of an electric power grid or distribution system. The computer 102 may be incorporated with the components of FIG. 12.
[0047] Step 101 of FIG. 1B includes monitoring and receiving observation data from a communication network 104. Although not shown in FIG. 1B, sensors integrated with the overhead distribution lines 112 are contemplated to detect irregular events and locations and provide observed data via a phasor management unit (PMU) connected to the communication network 104. The PMU (not shown) can provide the observation data to the grid operator for online decision support. The grid system can have PMUs positioned along the overhead distribution lines 112. The component fault prediction system can manage the PMUs and can have multiple components, i.e., processors, interfaces, etc., all of which may vary depending on the system configuration and user system and operational requirements.
[0048] 1A and 1B, a component failure prediction system 100B in FIG. 1B includes the steps of FIG. 1A and shows steps 111, 113, 117, 119 and 121 by the processor, and step 123 by the device, where steps not mentioned in FIG. 1B are assumed in FIG. 1A.
[0049] Some aspects of the components of the component failure prediction system 100B that are not shown in FIG. 1B can be optionally connected to a power grid system or an electric utility company's supervisory control and data acquisition (SCADA) system that monitors and relays information associated with the components, i.e., power line status and other infrastructure status. The specific configuration of components for an embodiment depends on user requirements and operational limitations, among other factors. With respect to SCADA, a user can set up the SCADA as an industrial control system or computer-based system for monitoring and controlling a process or equipment or infrastructure. Connected to the SCADA system can include a Human-Machine Interface (HMI), other computers and processors dedicated to certain tasks, and Remote Terminal Units (RTUs). The RTUs can connect to sensors, convert sensor signals to digital data, and send the digital data to a system of the SCADA system, which can be connected to a communication network 104 that communicates to the component failure prediction system 100B. Most control operations of the SCADA system can be performed by an RTU or by a programmable logic controller (PLC). While the PLC may control some processes, the SCADA system may allow the operator to change aspects of the processes of the component failure prediction system 100B. At the RTU or PLC level, this may include collecting data from other systems including equipment status reports that are communicated to the SCADA system via the communication network 104. The data is processed for the operator using an HMI, who makes decisions to adjust or change certain RTU and PLC controls. The databases and software programs of the SCADA system may be configured to provide trending and diagnostic data, as well as management information that can be communicated to the component failure prediction system 100B.
[0050] 1C is a schematic diagram illustrating a component failure prediction system 100C, similar to the method of FIG. 1A, showing a central computing application 131 connected to computer applications 133 and computing processing and data storage 135, according to some embodiments of the present disclosure. The central computing application 131 can include one or more computers capable of executing one or more applications 133, including an electric utility event incident management application (not shown). The central computing application 131 can include computing application processing and data storage 135, which can be incorporated with the system of FIG. 12 or one or more components of the system of FIG. 12, depending on the system and operational requirements of the user. The computing application processing and data storage 135 can include one or more computers 137, data storage 141 (i.e., can be cloud-based storage), local weather forecast service 151, real-time local weather observations 153, real-time component observations 155, real-time power circuit observations 157, and other observations 159, and can be directly connected to a communication network 132 that is connected to output information 134.
[0051] The failure component prediction model 137 may be integrated with or connected to one or more applications 133, namely, an electric utility storm incident management application (not shown). The failure component prediction model 137 may include an event incident crew 163 and an event other incident resource 165, all of which depend on user requirements and operational settings, as well as other user factors. In operation, an operator (not shown) may interface with one or more computing applications 133 via the central computing application 131, or via the failure component prediction model 137 via the central computing application 131, or both, depending on user operations and system requirements. Here, the operator may input information, view information, and generate data and information for the failure component prediction model 137 via the components shown in FIG. 12.
[0052] The embodiments contemplated by the one or more computational applications 133 may be embodiments that include applications having programs, some of which may include aspects having programs having instructions for modifying the forecast model to accommodate received observational data of local conditions measured in real time by a local network of weather monitoring stations that provide a comprehensive degree of resolution associated with the utility's supervisory control and data acquisition system, by selecting a parameterization scheme based on the received local conditions measured in real time by the local network and the difference between the forecasted conditions and the measured local conditions. Other programs may have instructions for presenting a user interface to allow user correction data to modify the results of the forecast model based on data conditions known by the operator and trend data that may be received as observed data or stored historical data, depending on when the particular knowledge is acquired (when an impending weather event is identified and prior to use of the model of the present disclosure). Still other programs may include instructions for initializing lateral boundary conditions based on larger scale numerical weather forecast data and lower boundary conditions based on climatological and geographic data of the domain of the forecast model, which may be received as observed data or stored historical data, depending on when the particular knowledge is acquired.
[0053] With respect to parameterization schemes, there may be cumulative boundary layer parameterization, surface layer parameterization, and microphysics, and each scheme may have sub-schemes. For example, an electric grid may require more weather condition information because it is concerned with damage to overhead power lines or some other components of the grid caused by weather variables including high winds and lightning. The electric grid may request an atmospheric science organization / agency to select a particular parameterization scheme optimized for some aspect of the convection, high winds, and lightning associated with it, which may be how the atmospheric science organization / agency parameterizes a particular feature in the model.
[0054] Parameterization in weather or climate models in the context of numerical weather prediction is a method of replacing processes that are too small or complex to be physically represented in the model by simplified processes. This is in contrast to other processes (e.g., large-scale flows of the atmosphere) that are explicitly solved in the model. Associated with these parameterizations are the various parameters used in the simplified processes. Examples include the falling speed of raindrops, convective clouds, simplifications of atmospheric radiative transfer based on atmospheric radiative transfer codes, and cloud microphysics. Radiative parameterization is equally important for both atmospheric and oceanic modeling. Atmospheric emissions from different sources within individual grid boxes also need to be parameterized to determine their impact on air quality.
[0055] 1C, the data storage 141 can include weather variables 143 corresponding to the model forcings (also referred to as a "set of weather drivers"), historical data 147, and other stored data 151. During operation, the failure component prediction model 137 can be connected to the data storage 141 to access the historical data 147 and other data 151. The historical data 147 can include weather data associated with the component data and the allocated resource data. Depending on the operational system configuration, the data can be generated based on past weather events, estimates of different types of resource allocations for an impending weather event at a given location. For example, some aspects of resource allocation for an impending weather event for a given location can include identifying resource allocation data corresponding to identified parameters of the impending weather event to estimate an output value. For example, the identified parameters of the impending weather event can include a corresponding set of data specific to resource allocation for a given location. Here, based on the availability of a data set of possible responses stored in data storage 141, the fault component prediction model 137 can estimate maintenance parameters such as, for example, predicted damage to the power circuit, predicted maintenance personnel-days to repair the damage, predicted consumer incidents from the damage, predicted estimated time to repair the power circuit, predicted estimated time to restore power to a particular customer, predicted estimated cost to repair the power circuit, etc.
[0056] It is contemplated that certain applications of the calculation applications 133 may also track real-time actual maintenance parameters, such as, for example, actual damage to the power circuit, actual maintenance personnel-days to repair the damage, actual consumer incidents from the damage, actual time to repair the power circuit, actual time to restore power to a particular customer, actual cost to repair the power circuit, etc., which are communicated to the failure prediction model 137 via the communications network 132.
[0057] 1C, it is contemplated that the communications network 132 may include one or more server computers (not shown) that communicate with client computers (not shown) for utilization by the electric utility storm incident management application (computer application 133). For example, the communications network 132 may be connected to client computers, cell phones, and other computing devices such as laptops and tablets. The communications network 132 may be a wireless network, a fixed line network, a local area network (LAN), a wide area network (WAN), an intranet, an extranet, the Internet, and the like. The communications network 132 may be the Internet, and for example, the server computer may be a web server with which the client computers communicate via any of a number of known communications protocols.
[0058] 1C, the local weather forecast service 151 may include forecasted wind speed and duration, forecasted storm duration, forecasted snowfall, forecasted ice amount, and forecasted rainfall, forecasted storm type (e.g., hurricane, wind, ice, tornado, lightning, etc.), forecasted lightning location and intensity, etc. The weather forecast may be embodied in or associated with a GIS (geographic information system) file, etc. The local weather forecast service 151 may include a national weather service office, a weather service organization, or some other weather forecast service. The real-time local weather observations 153 may be from sensors (not shown) located proximate to the predetermined location. The real-time components, i.e., overhead power lines, may generate sensor data via sensors (not shown) via component observations 155 located at the predetermined location. The real-time power circuit observations 157 may be from sensors (not shown) located proximate to the predetermined location. Other real-time observations 159 may include other data useful to the fault component prediction model 137 in generating results or data that may be associated with any aspect of the power grid system, including maintenance issues, operational issues, management issues, etc.
[0059] 2A is a schematic diagram illustrating some components of a power distribution system 200A and the relationships between some components of the system, according to some embodiments of the disclosure. The power distribution system 200A can be operated by a power distribution control system 201. The power distribution control system 201 can transmit power provided by a main power grid 210 through a substation 220 to one or more feeder breakers 240 to a power distribution line 234 load customer 205. The power distribution control system 201 can also include local power distribution generation sources 215, such as renewable generation and energy storage 217, through a conversion subsystem 216. The substation 220 can connect several feeders 240 through a bus 230. Each feeder 240 can include multiple power distribution line segments 235 and utility poles 255. Both the main power grid 210 and the power distribution control system 201 can be part of the main power grid 210. The distribution control system 201 may also be included in a control system (not shown) for a main distribution grid 210 .
[0060] 2A, the power distribution system 200A can adjust its topological connectivity by operating switching devices installed in the system, such as feeder breakers 240, normally closed switches or splitters 250, and normally open switches or tie switches 260. The power distribution system 200A can monitor its operational status by examining corresponding data collected from sensors 265 installed on the utility poles 255, segments 235, buses 230, and any other devices (not shown) associated with the operation of the power distribution system 200A. The sensors 265 can transmit measurements to the power distribution control system 201 via communication lines 270 or devices (not shown) mounted or located in proximity to other devices and systems related to the operation of the power distribution system 200A. The power distribution lines 234 and other devices (not shown) can become unavailable due to a fault at a fault location 246 for many reasons, including self-deterioration or environmental effects, such as a fallen tree 245 caused by a strong wind. In addition to sensor readings from sensors 265, distribution control system 201 may also monitor system abnormal events based on information provided by trouble calls from customers 205 or other sources.
[0061] FIG. 2B is a block diagram illustrating several method steps for an online power distribution line fault prediction problem, according to an embodiment of the present disclosure.
[0062]
number
[0063] According to an embodiment of the present disclosure, a tabular weather representation is first used to reduce the possible weather condition scenarios.The original problem is then transformed into multiple independent Bernoulli bandit problems.Finally, two types of learning frameworks are used to design a fault predictor.
[0064]
number
[0065] 2C is a schematic diagram illustrating a telephone company landscape 200C configured to predict the status of telephone components susceptible to severe weather events, according to an embodiment of the present disclosure. The telephone company 281 includes sensors 282 connected to sensor communication lines 286 (which may be wireless connections) and telephone lines 287 connected to utility poles 285. Residential customers 276 and business telephone customers 277 who receive telephone service are connected to the telephone lines 286. The telephone lines 286 include telephone line segments 283, and the utility poles include components 284. Similar to overhead distribution lines for the electric grid, the utility poles 285 have the same fault 279 problems with trees 278 and other vegetation close to the utility poles 285.
[0066] 2D is a schematic diagram illustrating a construction project for which some embodiments are configured to predict construction project component status susceptible to severe weather events, according to embodiments of the present disclosure. Construction project 200D includes components including formwork 291, masonry walls 292, walls 293, structural framing 294, and roof covering 295 that are susceptible to severe weather.
[0067] FIG. 3A is a schematic diagram illustrating a tabular weather representation according to an embodiment of the present disclosure.
[0068]
number
[0069]
number
[0070] If weather data of the required resolution for a given smaller area is not available, weather measurements for the larger area can be used to derive the required data based on the distance to the center of the larger area as measured from the smaller area under study. FIG. 3B is a schematic diagram illustrating deriving weather conditions based on measurements of a larger area according to an embodiment of the present disclosure.
[0071]
number
[0072] 4 is a block diagram illustrating some method steps of the transformed online power distribution line fault prediction problem according to an embodiment of the present disclosure. The weather predictor 407 provides weather forecasts to the weather renderer 405, which sends the processed weather conditions to the fault predictor 403, and the fault status of the power distribution line 401 is predicted.
[0073]
number
[0074]
number
[0075] We then use two learning frameworks to design fault predictors for the problem.
[0076]
number
[0077]
number
[0078]
number
[0079] 4, Algorithm 1 is a greedy action selection method that utilizes current knowledge to maximize immediate rewards without spending time exploring seemingly inferior actions and finding out whether they are actually better. To balance exploration and exploitation, this disclosure also provides an ε-greedy based estimation method and an upper confidence limit UCB based estimation method by making slight modifications to Algorithm 1.
[0080]
number
[0081] Still referring to FIG. 4, in the UCB method, the agent selects the action that has the highest estimated action value function plus a confidence upper bound search term, which takes into account the inherent uncertainty in the accuracy of its own estimates.
[0082]
number
[0083]
number
[0084]
number
[0085]
number
[0086]
number
[0087] With further reference to Figure 4, Algorithm 2 is a Thompson Sampling (TS) action selection method, which uses prior knowledge to explore the action space. As the posterior distribution becomes increasingly convergent, the agent explores less and exploits more, which strikes an effective balance.
[0088] To demonstrate the disclosed algorithm, an overhead distribution line is taken as an example. Two of the most influential weather events, namely wind W and lightning L, are selected as weather-related accident causes, and maximum instantaneous wind speed X and natural logarithm of lightning current Z are representative weather data, respectively. Consider two values of accident status state Y, Y=0 (i.e., in operation) and Y=1 (i.e., accident occurring). The relationship between W, L, X, Z, and Y can be expressed by a Bayesian network as shown in Figure 5A.
[0089] 5A is a schematic diagram illustrating a Bayesian network for data generation according to an embodiment of the present disclosure. It illustrates a data generation process with the following joint distributions:
[0090]
number
[0091] Based on (10), a dataset with n artificial data points can be generated by following the process specified in FIG. 5B.
[0092]
number
[0093]
number
[0094]
number
[0095] 6A and 6B are graphs illustrating aspects of a failure rate model and a failure probability model in Equation 11, according to an embodiment of the present disclosure.
[0096] 6A and 6B show the failure rate and failure probability in a one-to-one correspondence of points, i.e., the higher the failure rate, the higher the failure probability. The detailed parameter settings are shown in Table 1 of Fig. 7A for wind and Table 2 of Fig. 7B for lightning.
[0097] FIG. 7A is a table illustrating parameter settings for generating wind-related data according to an embodiment of the present disclosure, and FIG. 7B is a table illustrating parameter settings for generating lightning-related data according to an embodiment of the present disclosure.
[0098] Then, a dataset with 10,000 artificial data points is generated to verify the proposed online prediction algorithm. The data distribution is shown in Figure 8.
[0099] 8 is a graph showing data distribution of artificial data points, in which a data set having 10,000 artificial data points is generated to validate the online prediction algorithms described in Algorithm 1 and Algorithm 2 and their variants according to an embodiment of the present disclosure. It can be observed that most of the points are located in areas with small maximum instantaneous wind speeds and small lightning currents, but most of the accidents occur in weather events with large maximum instantaneous wind speeds or large lightning currents.
[0100] To evaluate the performance of the disclosed algorithms, the following two classification metrics are used:
[0101]
number
[0102] Here, TP is a true positive indicator, i.e., the algorithm predicts "accident occurring" and the actual status is actually "accident occurring", TN is a true negative indicator, i.e., the algorithm predicts "operational" and the actual status is actually "operational", FP is a false positive indicator, i.e., the algorithm predicts "accident occurring" but the actual status is "operational", and FN is a false negative indicator, i.e., the algorithm predicts "operational" but the actual status is "accident occurring".
[0103] 9A and 9B are graphs showing evaluation results using average precision and average F1 score, respectively, according to an embodiment of the present disclosure.
[0104] As shown in Figure 9A and Figure 9B, five action selection strategies are compared, where "mle" is the basic method that directly models the accident status using maximum likelihood estimation, "greedy", "ε-greedy" and "ucb" are action selection strategies based on Algorithm 1, and "ts" is based on Algorithm 2. As the time steps progress, each action selection strategy can converge, and "ts" can perform best with the highest average accuracy and F1 score after a period of learning. Other action selection strategies are not as good as "ts" because their exploration is not guided.
[0105] Furthermore, the density of the probability distribution for "ts" is also visualized in Figures 10A-10C and 11A-11C.
[0106] FIG. 10A is a graph illustrating the density of a probability distribution for the probability of a reward of 1 for giving a status prediction of 0 for a high occurrence weather event, according to an embodiment of the present disclosure.
[0107] FIG. 10B is a graph illustrating the density of a probability distribution for a probability of 1 of reward for giving a status prediction of 0 for a medium occurrence weather event, according to an embodiment of the present disclosure.
[0108] FIG. 10C is a graph illustrating the density of a probability distribution for a probability of 1 of reward for giving a status prediction of 0 for a low occurrence weather event, according to an embodiment of the present disclosure.
[0109] FIG. 11A is a graph illustrating the density of a probability distribution for probability 1 of reward for giving a status prediction 1 for a high occurrence weather event, according to an embodiment of the present disclosure.
[0110] FIG. 11B is a graph showing the density of a probability distribution for probability 1 of reward for giving a status prediction 1 for a medium-occurrence weather event according to an embodiment of the present disclosure.
[0111] FIG. 11C is a graph illustrating the density of a probability distribution for probability 1 of reward for giving a status prediction 1 for a low occurrence weather event according to an embodiment of the present disclosure.
[0112] Three representative weather blocks are selected: high occurrence weather events (Figures 10A and 11A), medium occurrence weather events (Figures 10B and 11B), and low occurrence weather events (Figures 10C and 11C). In some weather blocks, it can be observed that as the learning depth of the agent increases, the distribution of parameters concentrates around a certain point, which indicates a process of revising the data distribution through the feedback reward signal. This is also the reason why ts can perform better by constantly learning to find the most possible optimal parameters.
[0113] Tests with simulated data prove that both disclosed online fault prediction algorithms can effectively evaluate the impact of weather conditions on accidents in power distribution systems. In addition to being able to predict the accident status online with light computational complexity, these two methods can also utilize the reported accident data to update their knowledge online for the next decision-making task. Test results also show that the disclosed algorithms are effective for positive and negative samples, i.e., binary accident status. The Thompson sampling action selection strategy based on Algorithm 2 is the best compared to other strategies, and its embedded prior knowledge can aid the search for better learning performance.
[0114] Features One aspect of the disclosure may include the region being within a service area of an electric power system and including a set of components associated with an overhead distribution line of the electric power system. Another aspect is that the region is within a service area of an electric power system and the components are overhead distribution lines of the electric power system. The overhead distribution line includes (a) utility poles; (b) mechanical and electronic components; (c) distribution transformers mounted along the overhead distribution line utility poles; (d) protection devices; and (e) sensors installed on the overhead distribution line.
[0115] In another embodiment, the component data includes data received as the observational data and the stored data, the component data including, for each component, data including one or a combination of: age; type of material; specified type and design; mounting and support equipment; electrical equipment; vegetation assessment; pruning assessment; foliage assessment; installation and maintenance logs; severe weather exposure logs; assessment of vegetation and tree growth and type around overhead power lines; and management reports. In an embodiment, the weather data received as the observational data and the stored data includes data including one or a combination of: seasonal data; local area weather; snow, ice, rain / flood, wind, windstorm / lightning, ambient air temperature, wildfire, hurricane, severe weather assessment, severe windstorm assessment, classification of an imminent weather event as one of a mixed event, a pre-event, and a non-event.
[0116] According to an aspect of the disclosure, converting the set of weather variables into the table format includes a grid of weather cells for the domain, each weather cell representing a unique climate variability of the relationships between the weather variables in the set of weather variables. Another aspect is that the ML model for the set of weather drivers for a given component is configured by setting parameters of a predefined probability distribution for rewards of actions with initial values, the actions including predicting the component status as no fault and predicting the component status as fault. Yet another aspect is that the ML model for the set of weather drivers for a given component is updated through updating parameters of a probability distribution for rewards of actions and a corresponding action value function for each action, the actions including predicting the component status as no fault and predicting the component status as fault. Furthermore, an aspect is that a model forcing strength group is defined as a combination of variation intervals for the weather variables, each variation interval representing a unique variation range for one of the weather variables.
[0117]
number
[0118]
number
[0119]
number
[0120]
number
[0121]
number
[0122]
number
[0123]
number
[0124] Another aspect of the present disclosure can include a region that is a 2 mile grid, a 4 mile grid, or a 6 mile grid that includes components within an area served by an electric power system, comparing each weather cell to a weather variable threshold to identify a grouping of weather cells that meets the threshold that accurately represents a weather condition scenario in an area within the region.
[0125] Another aspect may include an application having programs, some of which may include aspects having programs having instructions for modifying a forecast model to correspond to local conditions measured in real time by a local network of weather monitoring stations providing a comprehensive degree of resolution associated with a utility's supervisory control and data acquisition system by selecting a parameterization scheme based on received local conditions measured in real time by the local network and differences between the measured forecast conditions and the local conditions. Another aspect may include a program having instructions for presenting a user interface for allowing user correction data to be modified by an operator based on known data conditions and trend data. Another aspect may include still other programs may include instructions for initializing lateral boundary conditions based on larger scale numerical weather forecast data and lower boundary conditions based on climatological-geographic data of the domain of the forecast model.
[0126] 1A and 1B is a block diagram illustrating the method of FIG. 1A and FIG. 1B, which can be implemented using alternative computers or hardware processors according to embodiments of the present disclosure. Computer 1211 includes hardware processor 1240, computer readable memory 1212, storage 1258, and user interface 1249 having display 1252 and keyboard 1251, connected via bus 1256. For example, user interface 1257, which communicates with hardware processor 1240 and computer readable memory 1212, obtains and stores signal data examples in computer readable memory 1212 when input from the surface of user interface 1257, keyboard surface 1257 by a user is received. 1140 The computer 1211 can include a power supply 1254, which, depending on the application, can optionally be located outside the computer 1211. Linked through the bus 1256 can be a user input interface 1257 adapted to connect to a display 1248, which can include, among other things, a computer monitor, a camera, a television, a projector, or a mobile device. A printer interface 1259 can also be connected via the bus 1256 and adapted to connect to a printing device 1232, which can include, among other things, a liquid inkjet printer, a solid inkjet printer, a large scale commercial printer, a thermal printer, a UV printer, or a dye sublimation printer. A network interface controller (NIC) 1254 is configured to connect to the network 1236 via the bus 1256 and can, among other things, render time series data or other data on a third-party display device, a third-party imaging device, and / or a third-party printing device external to the computer 1211.
[0127] 12 , the signal data or other data may be transmitted over a communication channel of the network 1236, among others, and / or stored in the storage system 1258 for storage and / or further processing. It is contemplated that the signal data may be initially stored in an external memory and later retrieved and processed by the hardware processor, or the signal data may be stored in the memory of the hardware processor for processing at some later time. The hardware processor memory includes stored executable programs executable by the hardware processor or computer to execute the component failure prediction system / method of the same type as the electrical distribution system, electrical distribution system operational data, and historical electrical distribution system data, as well as other data related to component failure prediction of the electrical distribution system or a similar type of electrical distribution system.
[0128] Additionally, signal data or other data may be received wirelessly or wired from receiver 1246 (or external receiver 1238) or transmitted wirelessly or wired via transmitter 1247 (or external transmitter 1239), both of which are connected via bus 1256. Computer 1211 may be connected to external sensing device 1244 and external input / output device 1241 via input interface 1208. For example, external sensing device 1244 may include sensors that collect data before, during, and after collected signal data of the power distribution system. For example, the faulted line segment, the fault time, and the customers affected by the fault. The computer 1211 may be connected to other external computers 1242. The output interface 1209 may be used to output the processed data from the hardware processor 1240. It should be noted that the user interface 1249, which communicates with the hardware processor 1240 and the non-transitory computer-readable storage medium 1212, receives input by a user from a surface 1252 of the user interface 1249, and obtains and stores the area data in the non-transitory computer-readable storage medium 1212. The bus 1256 may also connect other devices and interfaces, such as the GPS 1201, the printer interface 1259, to the memory device 1206 via the input interface 1208, and also to the external sensor 1204, and the machine 120.
[0129] definition Electricity transmission grid: and power grid refer to the same definition in this application. An electricity transmission grid is an interconnected network for distributing electricity from one location to another, e.g., from electricity producers to electricity consumers. An electricity transmission grid may consist of power plants that generate electricity, high voltage transmission lines that carry electricity from distant sources to demand centers, and distribution lines that connect to consumers of electricity, such as customers. For example, an electricity transmission grid or distribution grid may include one or more premises, such as homes, businesses, or facilities, that contain devices that consume electricity or power, and energy resources that provide electricity or power, such as generators and renewable energy.
[0130] Power Load: Can be an electrical load, an electrical component or part of a circuit that consumes (active) power. This is as opposed to a power source, such as a battery or generator, that produces the power. In a power circuit, examples of loads are appliances and lights. Loads can be further classified as critical and non-critical loads.
[0131] Critical Load: may require a substantially continuous supply of power and is therefore considered a "critical load". Some examples of critical loads include computers, computer-based control devices, and / or electronic data processing devices. With respect to critical loads and the supply of continuous power, a simple short interruption in the supply of standardized power by the utility grid may, for example, cause a computer to malfunction, which may sometimes be damaging and have serious consequences. Non-Critical Load: may be defined as any load that is not labeled as a critical load. Non-critical loads usually do not require a substantially continuous supply of power and may be shut off in the event of a generation shortage or restored after the critical load is restored. Fault Isolation: refers to the process by which faulty components are isolated from other healthy components after they have been identified and then, if possible, those components may be re-energized to restore power.
[0132] Operational Resiliency: can be defined as the ability of the electric distribution system as a whole to not contribute to, absorb and adapt to disasters. It goes beyond traditional operational risk and recovery capabilities, focusing on maintaining continuity of delivery of "critical functions" to both the electric distribution system and the system's customers.
[0133] Infrastructure resilience: Refers to the elasticity of hardware such as power towers, substation structures, and power lines.
[0134] Event: Any action that causes damage to at least a portion of the power grid, resulting in potential instability or loss of power in the power distribution network, which causes an interruption to the continuous power supply, either immediately or at some point in the near future. Some examples of events may be natural disaster events (weather, earthquakes, etc.), intentional damage events (terrorist attacks, etc.), or unintentional damage events (plane crashes, train crashes, etc.).
[0135] Restoration: Refers to the process of returning a partial or complete power distribution system to normal service after an extreme event has occurred that has caused a power outage to the partial or complete system. Service Restoration: Refers to the restoration of normal power supply to customers. Infrastructure Restoration: Refers to the restoration of damaged hardware such as utility poles, substation transformers, and power lines. Information Network: Refers to one of a wide area network, the Internet, or both.
[0136] Condition Information: Obtained from a device and may include device powered status, device damaged / disconnected status, terminal voltage, and power flow. For example, current condition information received from a device may be updated condition information for the moment the condition information is received or obtained.
[0137] Historical Data: refers to electrical grid data obtained before an extreme event occurs. For example, data may include location, time, type of weather event associated with component failures, weather and geographic data associated with reported failures and standard operation of each component, third party reported historical data associated with weather, geographic location, and terrain associated with incidents and not associated with incidents, maintenance, replacements, management past decisions, customer reported concerns during incidents and during normal operating hours without incidents. Any data associated with the operation of the electrical grid, including electrical grid components, weather, terrain, third party reported data, communication network status and reliability, etc.
[0138] Electric grid data: may include a topology of the electric grid incorporating identifying the location of the microgrid and one or more loads along with labeling the one or more loads as a subset of critical loads and a subset of non-critical loads.
[0139] Power Disruption: It may be a power outage or a power disturbance in an electric power distribution network. Some examples of causes of power disturbance may include a fault in a power generating station, damage to an electric transmission line of an electric power distribution system, a substation, or other parts of an electric power distribution system, a short circuit, or an overload of a power main. Specifically, a power outage may be a short or long term power loss condition in a given area or section of an electric power grid that may affect a single home, building, or an entire city, depending on the extent of the damage or cause of the outage.
[0140] Embodiment The following description provides exemplary embodiments only and is not intended to limit the scope, applicability, or configuration of the present disclosure. Rather, the following description of exemplary embodiments will provide one of ordinary skill in the art with an enabling description for implementing one or more exemplary embodiments. Contemplated are various changes that may be made in the function and arrangement of elements without departing from the spirit and scope of the disclosed subject matter as set forth in the claims.
[0141] Specific details are given in the following description to provide a thorough understanding of the embodiments. However, it will be understood by those skilled in the art that the embodiments may be practiced without these specific details. For example, systems, processes, and other elements in the disclosed subject matter may be shown as components in block diagram form so as not to obscure the embodiments in unnecessary detail. In other examples, well-known processes, structures, and techniques may be shown without unnecessary detail so as to avoid obscuring the embodiments. Furthermore, like reference numbers and names in the various drawings have indicated like elements. Also, individual embodiments may be described as a process that is shown as a flowchart, a flow diagram, a data flow diagram, a structure diagram, or a block diagram. Although a flowchart may describe operations as a sequential process, many of the operations may be performed in parallel or simultaneously. In addition, the order of operations may be permuted. A process may be terminated when its operations are completed, but may have additional steps not discussed or included in the diagram. Furthermore, not all operations in any specifically described process may occur in all embodiments. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, or the like. If a process corresponds to a function, the termination of the function may correspond to the return of the function to the calling function or to the main function.
[0142] Furthermore, embodiments of the disclosed subject matter may be implemented, at least in part, either manually or automatically. Manual or automatic implementations may be performed or at least assisted through the use of machines, hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, program code or code segments performing the necessary tasks may be stored in a machine-readable medium. A processor may perform the necessary tasks. The various methods or processes outlined herein may be coded as software executable on one or more processors using any one of a variety of operating systems or platforms. Furthermore, such software may be written using any of a number of suitable programming languages and / or programming or scripting tools, and compiled as executable machine code or intermediate code that runs on a framework or virtual machine. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments.
[0143] The embodiments of the present disclosure may be embodied as methods, of which an example is provided. The acts performed as part of the method may be ordered in any suitable manner. Thus, while shown in the exemplary embodiment as sequential operations, embodiments may be constructed in which operations are performed in an order different from that shown, which may include performing some operations simultaneously. Furthermore, the use of ordinal numbers such as "first," "second," and the like to modify claim elements in the claims does not, in itself, imply a priority, precedence, or order of one claim over another, or a chronological order in which the operations of the method are performed, but is merely used as a label to distinguish a claim element having a particular name from another element having the same name (but because of the use of ordinal numbers). Although the present disclosure has been described with reference to certain preferred embodiments, it will be understood that various other adaptations and modifications can be made within the spirit and scope of the present disclosure. It is therefore an aspect of the claims to encompass all such variations and modifications that fall within the true spirit and scope of the present disclosure.
Claims
1. A prediction system for a region, comprising: a component prediction processor, the component prediction processor comprising: receiving data from a data store including meteorological and component data for an area and generating a data set for the area to identify meteorological events with component locations, each of the meteorological events including a parameter, and a set of meteorological variables converted into a table format; generating a set of weather drivers, each including a given set of weather variables, and using the dataset to construct a machine learning (ML) model for each set of weather drivers for each component in the region; for each weather event in the region, iteratively for each component, identifying a corresponding set of weather drivers along with weather variables including maximum instantaneous wind speed and lightning current, generating output values for the ML model corresponding to the identified component for the set of weather variables, and updating the ML model with the output values and the component data for the weather event; receiving real-time observational data over a communication network of impending weather (IW) events in said region over a time period; and for each time period, repeatedly: Identifying, for each component, the same set of corresponding weather drivers along with weather variables, and updating the ML model corresponding to the identified component for the set of weather variables using the last run iteration forecast output values for the component and the observed data of the IW event; A prediction system configured to generate, for the updated ML model, an output value predicting a component status as failed or not failed for the time period.
2. The forecasting system of claim 1 , wherein the region is within a service area of an electric power system and the component is an overhead distribution line of the electric power system.
3. 3. The predictive system of claim 2, wherein the overhead distribution line includes: (a) utility poles; (b) mechanical and electronic components; (c) distribution transformers mounted along the utility poles of the overhead distribution line; (d) protection devices; and (e) sensors installed on the overhead distribution line.
4. 2. The forecasting system of claim 1, wherein the component data received as the observed and stored data includes, for each component, geographic data and data including one or a combination of the following: age; type of material; specified type and design; mounting and support equipment; electrical equipment; vegetation assessments; pruning assessments; foliage assessments; installation and maintenance logs; severe weather exposure logs; assessments of vegetation and tree growth and type around overhead power lines, and management reports.
5. 2. The forecasting system of claim 1, wherein the weather data received as observed data and stored data includes one or a combination of the following: seasonal data; local regional weather; classification of an impending weather event as one of: snow, ice, rain / flood, wind, windstorm / lightning, ambient air temperature, wildfire, hurricane, severe weather rating, severe windstorm rating, mixed event; non-event weather data, and the stored data includes pre- and post-weather events and non-event weather events.
6. 2. The forecasting system of claim 1, wherein converting the set of weather variables to the tabular form includes a grid of weather cells for the area, each weather cell representing a unique climate variability of relationships between the weather variables in the set of weather variables.
7. 2. The forecasting system of claim 1, wherein the ML model for a set of weather drivers for a given component is constructed by setting parameters of a predefined probability distribution for rewards of actions along with initial values, the actions including predicting a component status as failed or not failed.
8. 2. The forecasting system of claim 1, wherein the ML model for a set of weather drivers for a given component is updated through updating parameters of a probability distribution for rewards of actions and a corresponding action value function for each action, where actions include predicting a component status as failed or not failed.
9. 【Number 1】
10. 【Figure 2】
11. 【Number 3】
12. 【Fig. 4】
13. 【Number 5】
14. 【Number 6】
15. 【Number 7】
16. The forecasting system of claim 1 , further comprising: comparing each weather cell to weather variable thresholds to identify a grouping of weather cells that meets the thresholds that accurately represents a weather condition scenario in an area within the region.
17. A method for predicting components in a region, comprising: accessing data including meteorological and component data for a region and generating a data set for the region to identify meteorological events each including a parameter along with component locations, and a set of meteorological variables converted into a tabular format representing a grid of meteorological cells for the region; generating a set of weather drivers, each of which includes a given set of weather variables; constructing a machine learning (ML) model for each set of weather drivers for each entity in the region using the dataset; and iteratively: for each component, identifying a corresponding set of weather drivers along with weather variables including maximum instantaneous wind speed and lightning current, generating output values for the ML model corresponding to the identified component for the set of weather variables, and updating the ML model with the output values and the component data for the weather event; receiving real-time observational data for a period of time of an impending weather (IW) event in said region; and repeatedly, for each period: identifying, for each component, a corresponding set of meteorological drivers along with meteorological variables including maximum instantaneous wind speed and lightning current, and updating the ML model corresponding to the identified component for the set of meteorological variables using the last run iteration forecast output value for the component and the observed data of the IW event; and generating, for the updated ML model, an output value that predicts a component status as failed or not failed for the time period.
18. 1. One or more computer-readable storage media for predicting overhead power line status of an electrical grid system in a region, the one or more computer-readable storage media comprising: First program instructions; The first program instructions may include instructions for causing a computer to: (a) receiving data from a data store including meteorological and component data for an area and generating a data set for the area to identify meteorological events each including a parameter along with component locations, and a set of meteorological variables converted into a tabular format representing a grid of meteorological cells for the area; (b) generating a set of weather drivers, each including a given set of weather variables, and constructing a machine learning (ML) model for each set of weather drivers for each component in the region using the dataset; and iteratively: (c) for each component, identifying a corresponding set of meteorological drivers along with meteorological variables including maximum instantaneous wind speed and lightning current; (d) generating output values for the ML model corresponding to the identified components for the set of weather variables and updating the ML model with the output values and the component data for the weather event; The one or more computer-readable storage media further comprise second program instructions: The second program instructions may further include: (a) receiving real-time observational data over a period of time via a communication network of impending weather (IW) events in said region, and repeatedly, for each period of time: (b) identifying corresponding components and model forcing groups of the IW event from the observational data; (c) for each component, identifying a corresponding set of meteorological drivers along with meteorological variables including maximum instantaneous wind speed and lightning current, and updating the ML model corresponding to the identified component for the set of meteorological variables using the last run iteration forecast output value for the component and the observed data of the IW event; (d) generating, for the updated ML model, an output value that predicts a component status as failed or not failed for the time period.
19. A prediction system for components within a domain, comprising: The prediction processor receiving weather data and component data for a region from a memory; and generating a dataset for said region to identify parameters of weather events together with component locations, and generating a set of weather drivers, each of said set of weather drivers comprising a given set of weather variables converted into a tabular format representing a grid of weather cells configured to accurately represent weather condition scenarios in said region, said forecast processor further comprising: The data set is used to construct a machine learning (ML) model for each set of weather drivers for each component, and iteratively perform the following steps for each weather event: for each component, identifying a corresponding set of weather drivers along with weather variables, generating output values for the ML model corresponding to the identified component for the set of weather variables, and updating the ML model with the output values and the component data for the weather event; receiving real-time observational data over a period of time via a communication network of impending weather (IW) events in said region; and, repeatedly, for each period of time, A prediction system configured to: identify, for each component, a corresponding set of weather drivers along with weather variables; update the ML model corresponding to the identified component for the set of weather variables using a last run iteration forecast output value for the component and the observational data of the IW event; and generate, for the updated ML model, an output value predicting a component status as failed or not failed for the time period.
20. 20. The forecasting system of claim 19, wherein the components include one or a combination of: electric grid overhead distribution lines, electric grid components, utility poles, telephone components, or building project components including one or a combination of walls, formwork, roof coverings, and structural framing.
Citation Information
Patent Citations
Distributing system monitoring and controlling equipment
JP1997107634A
Forecast notification method and apparatus for thunder accident
JP2005181004A
US2014/025,7913A1
Storm response optimization
US20140257913A1
