3.7
Parameter Stochastic Models
In this section we present the stochastic models that we developed for our experiments. The focus is not to provide high accuracy models of the un- derlying processes, which in itself is a significant research task, but rather for them to be representative enough so that a reliable comparison of the EMS strategies can be made.
A particular requirement of the 2-stage EMS strategy is the generation of not just the expectation of the random parameters into the future, but also the generation of possible scenarios for these parameters conditioned on past events. We investigated a number of different model types before settling on generalised additive models for the continuous parameters like temperature, and Markov models for the more discrete occupant driven behaviours such as shiftable load requests.
3.7.1 Generalised Additive Models
A generalised additive model (GAM) [Hastie and Tibshirani, 1990] is a hybrid statistical model which combines properties of a generalised linear model (GLM) and an additive model (AM). For a GAM a probability dis- tribution from the exponential family is selected for the response random variableY. The expectation of the response is related to the predictor vari- ablesxi through a link function g that has been appropriately selected for
the chosen distribution:
g(E[Y]) =β0+f1(x1) +f2(x2) +. . . (3.34)
where the functionsfj are some parametric or non-parametric smooth func-
tions. The backfitting algorithm is one such algorithm that can be used for performing estimation.
The mgcv R package [Wood, 2011] was used for fitting the models and generating scenarios for the continuous stochastic parameters. Through ex- perimentation, we found a Gaussian distribution produced the best results for each model, which in effect reduces the model from a GAM to an AM. However, we make use of some of the more advanced features in the mgcv tool, such as using multivariate smooth functions, which are absent from the standard AM formulation.
We will now describe some of the models that were used for the continu- ous random parameters. The models were constructed to predict the value of a stochastic parameter rt at the upcoming time step, using the value of
the parameter at the previous time step rt−1, the time of the day tod, the
day of the week dow and other quantities as predictors.
Weather forecast data was used as predictors for the parameters temper- ature, global irradiance, wind and total network electricity demand (used to
estimate the electricity price). These forecasts and actual data for the para- meters relevant to the states of NSW and ACT in Australia were obtained from the Bureau of Meteorology (BOM) and the Australian Energy Market Operator (AEMO) and used to train the models and in EMS experiments. The forecasts include daily predictors of minimum and maximum temperat- ures tmin, tmax and morning and afternoon cloud covercldam, cldpm and wind speedswindam, windpm.
The mgcv models are presented below in mgcv pseudo-code for the con- tinuous stochastic parameters. The identify link function was used for each model, in line with the use of Gaussian distributions. Thesfunctions rep- resent smoothing functions which optionally have their basis dimension re- stricted by settingk.
E[tempt] = s(tempt−1) + s(tod,tmax,tmin,k = 20) (3.35)
E[globalt] = s(globalt−1) + s(tod,cldam,cldpm,k = 20) (3.36)
E[windt] = s(windt−1) + s(tod,windam,windpm,k = 20) (3.37)
E[demandt] = demandt−1+ s(tod,dow,tmax,k = 20) (3.38)
From top to bottom these are: the outdoor temperature, global irradiance, wind and aggregate network load. The wind and aggregate network load are used in predictions of the dynamic price as we will discuss shortly.
Once the models have been fit to training data, the expected values of the continuous variables over a forward horizon (for the distribution 3.25) can be calculated by repeatedly calling the mgcv predict function. Each time it is conditioned on the known predictors and previous values. In order to generate a scenario, a sample is drawn from a normal distribution with standard deviation equal to the result returned by the original fit, and then added to the result of the prediction step. This value is then used to condition the predictor for the next prediction, and the process is repeated until enough values for the number of time steps of interest have been generated.
Figure 3.4 shows five scenarios sampled from the outdoor temperature model. The temperatures are sampled for 10 days into the future, condi- tioned on the current temperature and the forecast minimum and maximum temperatures.
The best approach to implementing dynamic pricing in retail markets is still an open question. It will likely have a shape that is representative of the wholesale market, but with less volatility. We designed our dynamic price to be a quadratic function (which represents an increasing marginal supply price [Ramchurn et al., 2011]) of the amount of power that fossil fuel sources must supply to meet total network load. This is the total network demand (estimated with a GAM) minus the generation from renewable sources such as wind and solar. The dynamic price is only revealed to the EMS 30 minutes in advance.
3.7. PARAMETER STOCHASTIC MODELS 41
0
2
4
6
8
10
Time (days)
5
10
15
20
25
30
35
40
45
Temperature (degC)
Figure 3.4: Five scenarios for outdoor temperature, generated from the forecast minimums and maximums (dashed lines) over ten days.
The resulting prices have a similar magnitude to wholesale market prices in the NEM (see section 2.1). This is only a part of the overall residential retail prices, which also include network tariffs and other retail fees, but it is the part that has the strongest time dependence. This means that the costs presented in our experiments section should not be treated as absolute values as they do not include these additional retail charges and fixed costs.
3.7.2 Markov Models
A Markov model uses a random variable to model the state of the system, where the distribution of the variable in future states only depends on the current state, not on the states preceding it. In particular we make use of a semi-Markov models, which allows the distribution to also rely on the amount of time that the model has been in the current state.
Semi-Markov models were used to capture the behaviour of four occu- pants of a particular house in the ACT in Australia. These models provide the consumption patterns and profiles for the stochastic parameters such as hot water demand, uncontrollable energy consumption, shiftable load re- quests and EV usage. Each occupant had its own model that captured key activities, for example, watching TV, taking a shower and driving to work. Two dedicated models for clothes washing/drying (see figure 3.5) and dish washing were also developed. In addition to the states, the models also
specify the probability of transitioning from one activity to the next within certain time windows.
sleep dry morning dry evening sleep skip wash morning wash evening wash night dry night
Figure 3.5: Model of washing machine and clothes dryer usage.
A state transition can trigger the use of a resource. A simple example is thecooking activity/state which can trigger use of the oven resource. An- other one is thework activity which removes the EV resource from the home and turns the lighting off. These resources have a direct impact the para- meters in our optimisation problem. Conditional sampling through these models is used to generate scenarios.
Whilst this scheme was convenient for our experiments, other more data- driven options are possible: we could simply gather and use a database of raw scenarios, or learn model parameters from disaggregated demand data [Kolter et al., 2010, Parson et al., 2012].