The quantitative tools to support energy policy decisions range from assessment of macro-economic and cross-sectoral impacts (Kancs 2001; Siagian et al. 2017), to detailed micro-simulation models for a specific technology (Bhattacharyya 2011; Hunt and Evans 2009). Agent-based modeling (ABM) is a powerful tool for representing the complexities of energy demand, such as social interactions and spatial constraints and processes (Farmer and Foley 2009; Filatova et al. 2013). Unlike other approaches, ABM is not limited to perfectly rational agents or to abstract micro details in aggregate system-level equations. Instead, ABM can represent the behavior of energy consumers—such as individual households—using a range of behavioral theories. In addition, ABM has the ability to examine how interactions of heterogeneous agents at micro-level give rise to the emergence of macro outcomes, including those relevant for climate mitigation such as an adoption of low-carbon behavioral strategies and technologies over space and time (Rai and Henry 2016). The ABM approach simulates complex and nonlinear behavior that is intractable in equilibrium models.
However, this method is actively used in energy applications to study national climate mitigation strategies (Gerst et al. 2013; Gotts and Polhill 2017), energy producer behavior (Aliabadi et al. 2017), renewable energy auctions (Anatolitis and Welisch 2017), consumer adoption of energy-efficient technology (Chappin and Afman 2013; Jackson 2010; Palmer et al. 2015; Rai and Robinson 2015), shifts in consumption patterns (Bravo et al. 2013), changes in energy policy processes (Iychettira et al. 2017), and diffusion of energy-related actions and technology (Ernst and Briegel 2017; Kangur et al. 2017). Many cases of ABM still either lack a theoretical framework (Groeneveld et al. 2017) or relevance to empirical data, especially when studying energy behavior of households (Amouroux et al. 2013).
To assess the impact of individual behavior on carbon emissions, we went beyond classical economic models and the stylized representation of a perfectly informed optimizer. Therefore, we further developed the BENCHFootnote 6 agent-based model (Niamir et al. 2018a) by strengthening the alignment of behavioral and economic factors under different climate policy scenarios. We calibrated the BENCH-v.2 model using data on households’ energy-related choices from a survey specially designed for this purpose (Sect. 2.3) and administered in a European region of Overijssel, The Netherlands (1383 households). The BENCH-v.2 calculates changes in electricity consumption annually and implied carbon emission—based on the primary source of energy—by simulating individuals’ behaviors (Sect. 3).
Overview: individual energy behavior
There is a number of energy-related actions in which individuals may pursue to influence their electricity consumption and, consequently, their carbon footprint. We categorize them into three main types of behavioral changes. An individual can make an investment (action A1), either large (such as installing solar panels) or small (such as buying energy-efficient appliances, e.g., A++ washing machine). Alternatively, individuals can save energy by changing their daily routines and habits (action A2)—e.g. by switching off the extra lights and adjusting a thermostat/air conditioner. Finally, households can switch to a supplier that provides green electricity (action A3) (Niamir and Filatova 2017).
A decision is a process through which the selection of one among numerous possible behavior alternatives is performed (Barros 2010; Simon et al. 1997). Individuals are often bounded by their own previous experiences and their cognitive abilities—personal aspect—the influence of others—social aspect—and information availability. Empirical studies in psychology and behavioral economics show that individual choices and behaviors often deviate from the assumptions of rationality: there are persistent biases in human decision-making (Frederiks et al. 2015; Kahneman 2003; Niamir and Filatova 2016; Pollitt and Shaorshadze 2013; Stern 2013; Wilson and Dowlatabadi 2007). Driven by the empirical evidence from environmental behavioral studies (Abrahamse and Steg 2011; Bamberg et al. 2007; Bamberg et al. 2015; Mills and Schleich 2012; Onwezen et al. 2013; Steg and Vlek 2009), the BENCH-v.2 model assumes that a decision regarding any of the three actions (A1–A3) is driven by psychological and social factors in addition to the standard economic drivers such as prices relative to incomes (Niamir et al. 2018a). Behavioral factors including personal norms and awareness may either amplify the economic logic behind a decision-making or impede it, serving either as a trigger or a barrier. It is a scientific challenge to combine the behavioral and the economic parts of the decision-making process in a formal model. Here, we present the simplest option assigning weights to the behavioral part by calculating households’ intentions toward a specific energy-related action derived from our household survey dataset.
Survey and empirical data
Our household survey is designed to elicit factors and stages of a decision-making process with respect to the three types of actions that households typically make (A1 investment, A2 conservation, and A3 switching). The conceptual framework behind the survey assumes three main steps that lead to one of these actions: knowledge activation, motivation, and consideration (Niamir et al. 2018a). Before considering action, households need to reach a certain level of knowledge and awareness about climate change, energy, and the environment. If an individual in a household is aware enough, she might feel guiltFootnote 7. Here, personal norms (individual attitudes and beliefs) and subjective norms prevailing in a society add to her motivation. If households get motivated, they feel responsible to do something. Still, none of these factors are enough to provoke an action to change the energy use behavior. A household needs to consider its economic status, its house conditions (e.g. renting of owning), its current habits, and own perception of its ability to perform an action or change behavior. If a household reaches a certain level of intention, it is going to decide or act.
To elicit data on an interplay of behavioral and economic factors, we conducted a survey in a European region (NUTS2 level) in 2016: Overijssel province in The Netherlands (NL21), see Appendix, Fig. A2. The data on the behavioral and economic factors affecting household energy choices were collected using an online questionnaire (N = 1383 households in Overijssel) and serve as empirical micro-foundation of agent rules in the BENCH-v.2 model. The variations in socio-demographic and psychological factors among the respondents are further used to initialize a population of heterogeneous agents in the ABM (Sect. 2.3). The differentiation per income group also allows to potentially connect with other micro and macro statistical data if needed.
BENCH agent-based model
Compared to its first version (Niamir et al. 2018a), the BENCH ABM has been further developed and modified to investigate the macro impact of cumulative individual behavioral change on carbon emissions. In particular, in this application, we extended BENCH by (a) introducing three representative electricity producers (gray, brown, and green); (b) further improving the model engine, which now treats behavioral and economical parts explicitly (Sect. 2.1). In the behavioral part, the psychological and social aspects of a household’s behavior change and decision making are evaluated (Sect. 2.3.1). If there is high intention, household agents proceed with assessing the typical economic utility (Sect. 2.3.2). We combine and harmonize the behavioral and the economic parts of the decision-making process by extending the standard utility function (Eq. 3, Sect. 2.3.2). Here, an individual may overcome her economic barrier, if the behavioral part outweighs, e.g., the level of knowledge, motivation, and intention raise high enough to reconsider the economic tradeoffs. It goes in line with empirical findings revealing that individual willingness to pay for renewable energies, e.g., green electricity, is beyond the economic concept and monetary pay-off (Lee and Heo 2016; Sundt and Rehdanz 2015). In the economic part, households’ utilities based on the three actions (A1–A3) are calculated and compared (Fig. 1).
Further changes compared to the original BENCH include (c) improvements in social dynamics and learning algorithms by introducing and simulating two ways of households’ interactions (Sect. 2.3.3); (d) running a carbon price scenario as a top-down strategy to investigate impacts of policies on household behavioral change (Sect. 2.3.4); (e) the results of simulations in terms of CO2 emissions (tons per capita) to compare between scenarios (Sect. 2.4, 3) to get a better overview on the impacts of individuals’ behavior on carbon emissions over time and space. The role of each action (A1–A3) in these trajectories is also estimated till 2030 (Sect. 3).
Household agents in BENCH-v.2 are heterogeneous in socio-economic characteristics, preferences, and awareness of environment and climate change, so they can pursue various energy-related choices and actions. Namely, they vary in six economic attributes: (1) annual income in euro; (2) annual electricity consumption in kWh; (3) household status in terms of being a gray, brown, or green electricity user; (4) dwelling tenure status—owner or renter; (5) energy label of their dwelling varying from A to F; and (6) the household energy use routines and habits measured in the survey in terms of frequency of performing a particular energy-consuming action. Data for all these variables come from the survey. The annual growth value of socio-economic variables representing households’ income, electricity consumption, and consumption of other goods (in 5 quintiles) for the Overijssel province comes from the EXIOMODFootnote 8 computable general equilibrium (CGE) model (Belete et al. 2019). The behavioral and social aspects impacting households energy decisions also vary among agents and include (1) personal normsFootnote 9, which are values that people hold (Schwartz 1977), e.g., feeling good when using energy-efficient equipment; (2) subjective normsFootnote 10, which are perceived social pressure on whether to engage in a specific behavior motivated by observing energy-related actions of neighbors, family, and friends; and (3) perceived behavioral control (Sect. 2.3.1). These behavioral and social variables are updated over time (annually) through social dynamics and learning procedures (Sect. 2.3.3). Agents’ decision processes closely follow the conceptual framework (Fig. 1) behind the household survey and apply to all three types of energy-related behaviors (A1–A3).
Behavior part
Based on different internal and external barriers and drivers, households have different knowledge and awareness levels about the state of the climate and environment, motivation levels to change their energy behavior, and consideration levels when they perform costs and utility assessments. All household attributes are heterogeneous and change over time and space. All the variables in knowledge activation, motivation, and consideration are measured in comparable ways using Likert scale, in the range of 1–7 as in the survey. Here, 1 stands for the lowest, 7 is the highest level (Niamir et al. 2017).
Niamir et al. (2018a) described how households’ knowledge and awareness (K) and motivation (Mn) are measured and calculated at the model initialization stage based on the survey data. In summary, K is based on climate-energy-environment knowledge (CEEK), climate-energy-environment awareness (CEEA), and energy-related decision awareness (EDA) values. If households are aware enough, that is they have a high level of knowledge and awareness above the threshold of 5 out of 7, then they are tagged as “feeling guilt” and proceed to the next step to assess their motivation (Mn) for particular actions. Households’ personal norms (PNn) and subjective norms (SNn) are assessed to calculate their motivation (Mn). In this paper, motivation may differ for each of the three main actions (n = {1,2,3}). For example, a household may have a high level of motivation for installing solar panels, and is therefore tagged as “responsible” for action 1 (investment) and proceeds to the next step (consideration). At the same time, it may not pass the threshold value in motivation for changing energy use habits or switching to another energy supplier, and thus does not go into the consideration step on those two actions. If household agents have a high motivation level and feel responsible, they consider the psychological (e.g., perceived behavior controlFootnote 11), structural (housing attributes), and institutional factors (e.g., subsidies) to assess utility and costs of a specific action (Sect. 2.3.2). Then, households with high level of consideration are tagged as “high intention”. In the consideration stage, as well as the motivation stage, we differentiate between actions. In investment (A1) for instance, the dwelling ownership status (SF, owner or renter) and perceived behavioral control over the investment (PBC1) are checked and evaluated (δ1). While the ownership status is not essential in conservation (A2) and switching (A3), δ2 and δ3 are calculated just based on perceived behavioral controls (PBC2 and PBC3). All this is captured by the following equations:
$$ {\displaystyle \begin{array}{c}\begin{array}{c}K=\frac{AVG\ \left( CEEK, CEEA, EDA\right)}{7};\\ {}{M}_n=\frac{AVG\ \left({PN}_n,{SN}_n\right)}{7};\end{array}\\ {} If\ \left(n=1\ and\ SF=1\right)\ \left(\ {\delta}_1=\frac{PBC_1}{7}\ \right)\ else\ \left(\ {\delta}_1=0\right);\\ {} If\ \left(n=2\ \right)\ \left(\ {\delta}_2=\frac{PBC_2}{7}\ \right); If\ \left(n=3\ \right)\ \left(\ {\delta}_3=\frac{PBC_3}{7}\ \right)\end{array}} $$
(1)
Economic part
The economic part estimates utility of an individual agent for undertaking any of the three main actions. Energy economics (Bhattacharyya 2011) assumes that households receive utility from consuming energy (E, here green, brown, or gray) and a composite good (Z) under budget constraints:
$$ U=Z\cdotp \alpha +E\cdotp \left(1-\alpha \right) $$
(2)
Here, α is the share of individual annual income spent on the composite good.
Niamir et al. (2018a) extend this standard utility by including the influence of knowledge and awareness (K) and motivation (Mn) and adding actions’ intention (δn) as a weight on the behavioral part:
$$ U=\left(Z\cdotp \alpha +E\cdotp \left(1-\alpha \right)\right)\cdotp \left(1-{\delta}_n\right)+\left(K+ Mn\right).{\delta}_n $$
(3)
This weight is calculated and normalized using the survey data.
Social dynamics and learning
Heterogeneous households engage in interactions and learn from each other. In particular, they can exchange information with neighbors, which may alter own knowledge, awareness, and motivation regarding energy-related behavior. Here, we employ a simple opinion dynamics model (Acemoglu and Ozdaglar 2011; Degroot 1974; Hegselmann and Krause 2002; Moussaid et al. 2015) assuming that each agent interacts with a fixed set of nearby neighbors. Agents compare values of their own behavioral factors—knowledge, awareness, and motivation—with those of their eight closest neighbors, and adjust their values for a closer match. In different scenarios (Table 1), we introduce two types of interaction dynamics among households: slow and fast. Following the slow dynamics, households in an active neighborhoodFootnote 12 interact with maximally two neighbors (households 3 and 4 in Fig. 2a), and a household(s) with lower than average value of the whole neighborhood increases their current value by 5% (Fig. 2a). In the fast dynamics configuration, all households in an active neighborhood exchange of opinions and learn from each other (Fig. 2b, Eq. 4). In addition, the related perceived behavior control (PBCn) of a household that already took an action (household 5 in Fig. 2) is raised by 5% (Eq. 5). Future research may focus on advancing this social dynamics further, by for example differentiating per type of energy-efficiency action (observable or not) or dynamics of diffusion process. Moreover, different channels to establish a social network may be relevant for individual decisions. Understanding how the structure of social networks initiated based on friendship, family, and other relationships beyond the spatial distance (Allcott 2011; Jachimowicz et al. 2018) alone is a prominent future research direction, potentially supported by big data form social media. Similarly, future research may focus on assessing the consequences of social network structures—regular, small-world, or scale-free networks (Newman, 2003; Watts, 2004)—on the aggregated energy and CO2 emission dynamics.
Table 1 End-user scenario settings: climate policy and human behavior scenarios
$$ {\displaystyle \begin{array}{c}X=\left\{ CEEK, CEEA, EDA,{PN}_n,{SN}_n\right\},n=\left\{1,\dots 9\right\}\kern0.5em ;\\ {} If\ \mathit{\operatorname{Max}}\ \left( mean\ \left({X}_n^t\right), median\ \left({X}_n^t\right)\right)\ge {X}_3^t\kern1.25em \left({X}_3^{t+1}={X}_3^t+0.05\cdotp {X}_3^t\right)\kern0.75em ;\\ {} If\ \mathit{\operatorname{Max}}\ \left( mean\ \left({X}_n^t\right), median\ \left({X}_n^t\right)\right)\ge {X}_4^t\kern1.25em \left({X}_4^{t+1}={X}_4^t+0.05\cdotp {X}_4^t\right)\end{array}} $$
(4)
$$ {PBC}_5^{t+1}=\kern0.5em {PBC}_5^t+0.05\cdotp {PBC}_5^t; $$
(5)
Carbon emissions and pricing
In this research, we investigate CO2 emissions implied by households’ electricity consumption which is supplied from power plants using different kinds of fuels. Carbon dioxide emission factors for electricity have been derived as the ratio of CO2 emissions from fuel inputs of power plants relative to the electricity delivered. CO2 emission factors of each fuel type are used as defined in IPCC (2006). Three different kinds of electricity suppliers are considered, between which the households can choose: “gray”, “brown”, and “green”. The assumptions regarding fuel mixes and the resulting net CO2 emission factors are listed in Appendix, Table A1.
To estimate the impact of climate policies, namely a carbon price, we design and add climate policy scenarios by including carbon price in the utility estimations of households.
End-user scenarios
Traditionally, rational optimization models such as CGE models, have been used to predict household energy consumption under various socio-economic scenarios including shared socioeconomic pathways (SSP)Footnote 13. Here, the baseline scenario represents this traditional economic setup where rational and fully informed households make optimal decisions. Therefore, we use aggregated residential electricity consumption from the EXIMOD model downscaled to the regional level. The baseline scenario (gray dash-line in Figs. 3 and 5) is an output of this CGE model under SSP2 (business as usual).
We use this baseline scenario as a benchmark to compare the output of our behaviorally rich ABM. Four end-user scenarios in BENCH.v2 are designed to explore the impacts of heterogeneity in household attributes such as income and electricity consumption, social dynamics (bottom-up approach), and carbon price pressure (top-down approach) strategies on the individual and aggregated household behavioral change (Table 1). In all cases, based on the energy behavior change of households, we assess the following macro-metrics at the regional level: the diffusion of each of the three types of behavioral actions (A1–3) among households over time, and the changes in carbon emission reduction per capita.