Modelling of sanitary hot water energy consumption using adaptive neuro-fuzzy inference systems

[摘要] ENGLISH ABSTRACT: The introduction of Energy Management (EM) interventions aimed at reducing electrical energy consumption in the residential, commercial and industrial load sectors has expanded rapidly on a global scale in recent years. These programs are driven by environmental concerns, capacity constraints experienced in generation and transmission and the need to improve end-use efficiency. The implementation of EM schemes often involves financial incentives funded by governments and utilities. Measurement and Verification (M&V) performance assessments aimed at determining the savings impacts form an integral part of the management of these EM incentive programmes. M&V baseline development involves the development and implementation of accurate models that relate the energy consumption of a targeted load to variable energy-governing factors in order to determine actual savings impacts.The electrical energy consumption associated with sanitary water heating represents a sizeable component of the cumulative energy consumption associated with a number of load categories found in the various load sectors. In general, the electricity consumption profiles associated with sanitary hot water consumption correlates closely with the household electricity consumption profiles found in the residential load sector, particularly in the sense that it is influenced by the same socio-economic factors and human behavioural patterns. Soft computing methods have been employed successfully for residential load prediction, as these are tolerant of stochastic behaviour and uncertainty and do not require exact input to output matching. Particular success in the field of residential Short Term Load Forecasting (STLF) has been achieved using Adaptive Neuro-Fuzzy Inference Systems (ANFIS).An ANFIS load forecasting model with a long prediction horizon of up to a year is found to be capable of reasonable modelling accuracy for the estimation of the time-series profile of a system. It also exhibits very good prediction accuracy when calculating the total energy use over time of that profile. The load data used in this study is of student residence heat pump power consumption profiles and spans over four years with 48 samples for each day. The training inputs that are considered other than the load are the time of the day, the day of the week, the day of the year and the temperature.After the proof of concept, a comparative case study is performed with the view to explore optimal configurations of differing inputs to the ANFIS method. The effects of compartmentalising the dataset into subsets representing different characteristics, thereby deriving different models representing different cyclic periods, are also explored. It is found that compartmentalising the load model into 48 ANFIS sub-models, each serving a specific half-hourly time period in the day, results in the most best modelling accuracy.

[发布机构] Stellenbosch University