Predictive Analytics build systems models from historical operational data. Over time, the accumulated Industrial IoT data can add up to very big data sets. Our Predictive Analytics extend baseline models that are based on engineering knowledge and scale them up to model asset fleets.
The baseline models may incorporate existing knowledge and come in the form of physics-based models, regression models, or other machine learning models. Mitek offers advanced analytics that scale the baseline models, improve confidence, and create new functionality. As an example, a model that describes performance of one asset at one time period can be scaled up across the population and along many time periods to provide insights about cross-fleet variation and trends. As another example, a forecasting model that predicts the expected value of demand can be scaled to a non-parametric model of entire multivariable probability distribution of the demad; such distribution model is useful for stochastic optimization of supply decisions.
Our Predictive Analytics software uses innovative mathematical algorithms for Big Data batch processing. The algorithms perform optimal Bayesian estimation of model parameters. Mitek’s tools address both parametric performance data and discrete event data. The model parameters are obtained by using convex optimization methods for fast, precise, and scalable (parallelizable) computation of the results.
Our analytics include advanced automated data cleanup, which is absolutely necessary for any large scale advanced analytics. Typical big data sets in the Industrial IoT domain might have data problems in 10%, or more, records. At the same time, even a few unnoticed outliers might bias model estimation and cause loss of the model fidelity.
Asset Fleet Analytics
Mitek Analytics provides unique analytics technologies for multi-level analysis of the performance data for fleets of IIoT assets. The analysis includes variability across the population and trends along the time. In one project, a multi-level model with 10 million regressors has been estimated from 1 trillion data points (flight data for 300-aircraft fleet over 3 years); these computations take 12 hours on a 20-core cluster. Another example is estimation of reliability variations and trends across the population of several thousand reparable parts from multi-year maintenance and repair record data; the aggregated event data are processed in several seconds.
The models for asset fleets can be build based on historical operational data. This allows to deploy asset fleet analytics even when the asset original manufacturer does not provide suitable models.
Probabilistic Forecasting Analytics
Mitek Analytics provides unique analytics technologies for data-driven probabilistic modeling of forecasts. These data-driven models enable practical implementation of probabilistically optimized business decisions. One application is to electricity markets where errors of forecasting electrical power loads might lead to undesirable outcomes. The problems of balancing power grid are important because of the variability brought by increasing renewable generation. As one example, 2-3% cost reduction compared to the state of the art in the day-ahead electricty markets is possible by using empirical data-driven probabilistic models that describe joint distribution of the power load and power prices.
Another exampel is analysis of the generating capacity adequacy in planning of electrical grid expansion. This requires predictive analysis of extreme combinations of the loads, generation outages, and renewable generation. Our probabilistic modeling includes the tails of the distributions that represent extreme events such as peak power loads, or extreme forecast errors, or extreme weather events. Mitek works with power system operators to provide data-driven probabilistic risk assessment for adequacy analysis.