On Price Responsive Consumer Behavior in Electricity Markets: to Machina Economicus from Homo Agens

The electricity power market is well known for its highly volatile nature due to its innate variability characteristic of demand and the absence of practical bulk storage at reasonable cost. Any discordance between rapid fluctuation in wholesale prices and near flat retail prices not only incurs economic inefficiency in terms of social welfare, but also creates price-inelastic wholesale demand which severely exacerbates the volatility of wholesale electricity prices. While the market has a fundamental dynamic nature, the behavioral aspect of power consumption in response to price changes is not well understood. This necessitate to develop a empirical modeling methodology of demand which can potentially provide practical insights into demand response. In the former part of this work, we focus on dynamic aspect of demand response in Chapter 2. We first show that (i) demand is well responsive to outlier high price surges, and (ii) demand response can incur a certain amount of delay. Examining further data, it appears that demand is responsive to anticipated prices. This is in conformity with our previous observations on the inertia of demand, and testing the hypothesis that demand actually responds to anticipated prices rather than actual real time prices is an important next step. While it is impractical to obtain a particular individual’s own price prediction, We propose to test the hypothesis with day-ahead electricity prices (DAP). In addition, as an initial step toward the derivation of a quantitative model of electricity load and price, we propose a model of “appliance” usage as a representative basic component of electricity load. In the latter part of this work, we investigate more fundamental aspect of data-centric modeling in Chapter 3. First, we show the limitation of pure data-centric modeling strategy by proving that having a perfect knowledge on the joint distribution on price and load does not identify the load behavior in response to price. As it turns out that the causal structure of the variables of interest is the central matter that determines load behavior identifiability, we derive a minimal identifiable causal structure of demand response from the preexisting economic theories. Based on the discovered causal structure, we propose a minimal Bayesian model representation called “stochastic neuron” which connects machine learning technique to demand response modeling. We show that a stochastic neuron is an explainable tool as expressive as an ordinary neural network, and well extends the arguments from “appliance” usage model