Оптимизация прогноза несиационарных объектов на основе интеллектуфлного регулирования значений переменных
- № 3(15) 2018
Страницы:
111
–
126
Язык: русский
Аннотация
Сформулирована проблема оптимизации прогноза на базисе концептуальных принципов и методов многокомпонентного моделирования нестационарных объектов и синтеза вычислительных схем динамических моделей, нечетких множеств, нечеткой логики и нейронных сетей (НС). Предложен гибридный подход, основанный на методы поиска глобального и локальных экстремумов при оптимизации идентификации и обработки данных для обеспечения достоверности прогноза случайных временны хрядов (СВР). Разработан метод обеспечения достоверности прогноза СВР на основе применения нейро-нечетких сетей (ННС), реализованы модифицированные вычислительные схемы структурной и параметрической идентификации, нечеткой логики, применения базы знаний (БЗ), включающих широкий на борнечетких правил взамен сложных аналитических функций и уравнений, описывающих нестационарный процесс. Реализована вычислительная схема нечеткого вывода Сугено нулевого порядка и пятислойной НС, выполняющие функции формирования терм входных переменных, антецедентов нечетких правил, нормализации степеней выполнения правил, формирования заключений по нечетким правилам, агрегирования результатов по различным правилам. Разработан алгоритм обучения НС с определением и настройкой параметров функций принадлежностей (ФП) носителя нечеткихмножеств,которыйосновываетсянаалгоритмсобратнымраспространением ошибки по методам наименьших квадратов и градиентной оптимизации. Метод идентификации СВР совершенствован путем синтеза алгоритмов полиномиального нелинейного фильтра, нечетких множеств, регулирования переменных на основе нечеткой логики и нейронной сети. Разработанные алгоритмы идентификации СВР, оптимизации, регулирования и обработки данных реализованы в составе программного комплекса прогнозирования СВР и проведен сравнительный анализ их эффективности.
The problem is formulated for forecast optimization on the basis of conceptual principles and techniques of multicomponent modeling of non-stationary objects and synthesis of computing schemes of dynamic models, fuzzy sets, fuzzy logic and neural networks (NN). A hybrid approach based on methods of searching for global and local extremes during optimization of identification and processing of data to ensure the reliability of forecast of random time series (RTS) is proposed. Methods is developed to ensure the reliability of RTS forecast on the basis of application of neural-fuzzy networks (NFN), modified computational schemes of structural and parametric identification, fuzzy logic, using of knowledge base (KB) including a wide set of fuzzy rules instead of complex analytical functions and equations describing nonstationary process. Implemented computational scheme of fuzzy inference Sugeno of zero level and five-layered NN have functions of forming the term of input variables, antecedents of fuzzy rules, normalization degrees rule execution, forming conclusions of fuzzy rules, aggregation of results obtained by different rules. NN training algorithm is designed for defining and setting the parameters of belonging functions (BF) of fuzzy sets carrier, which is based on an algorithm to reverse the spread of least squares and gradient optimization mistakes. The method of OTS identification is improved by synthesizing the algorithms of polynomial nonlinear filter, fuzzy sets, regulation of variables based on fuzzy logic and neural network. The developed algorithms for OTS identification, data optimization, regulation and processing were implemented as part of the software complex for OTS forecasting and comparative analysis of their effectiveness are carried out.
The problem is formulated for forecast optimization on the basis of conceptual principles and techniques of multicomponent modeling of non-stationary objects and synthesis of computing schemes of dynamic models, fuzzy sets, fuzzy logic and neural networks (NN). A hybrid approach based on methods of searching for global and local extremes during optimization of identification and processing of data to ensure the reliability of forecast of random time series (RTS) is proposed. Methods is developed to ensure the reliability of RTS forecast on the basis of application of neural-fuzzy networks (NFN), modified computational schemes of structural and parametric identification, fuzzy logic, using of knowledge base (KB) including a wide set of fuzzy rules instead of complex analytical functions and equations describing nonstationary process. Implemented computational scheme of fuzzy inference Sugeno of zero level and five-layered NN have functions of forming the term of input variables, antecedents of fuzzy rules, normalization degrees rule execution, forming conclusions of fuzzy rules, aggregation of results obtained by different rules. NN training algorithm is designed for defining and setting the parameters of belonging functions (BF) of fuzzy sets carrier, which is based on an algorithm to reverse the spread of least squares and gradient optimization mistakes. The method of OTS identification is improved by synthesizing the algorithms of polynomial nonlinear filter, fuzzy sets, regulation of variables based on fuzzy logic and neural network. The developed algorithms for OTS identification, data optimization, regulation and processing were implemented as part of the software complex for OTS forecasting and comparative analysis of their effectiveness are carried out.