Can Artificial Intelligence Assist Project Developers in Long-Term Management of Energy Projects? The Case of CO₂ Capture and Storage

Abstract

This paper contributes to the state of the art of applications of artificial intelligence (AI) in energy systems with a focus on the phenomenon of social acceptance of energy projects. The aim of the paper is to present a novel AI-powered communication and engagement framework for energy projects. The method can assist project managers of energy projects to develop AI-powered virtual communication and engagement agents for engaging their citizens and their network of stakeholders who influence their energy projects. Unlike the standard consultation techniques and large-scale deliberative engagement approaches that require face-to-face engagement, the virtual engagement platform provides citizens with a forum to continually influence project outcomes at the comfort of their homes or anywhere via mobile devices. In the communication and engagement process, the project managers’ cognitive capability can be augmented with the probabilistic capability of the algorithm to gain insights into the stakeholders’ positive and negative feelings on the project, in order to devise interventions to co-develop an acceptable energy project. The proposed method was developed using the combined capability of fuzzy logic and a deep neural network incorporated with a Likert scaling strategy to reason with and engage people. In a mainstream deep neural network, one requires lots of data to build the system. The novelty of our system, however, in relation to the mainstream deep neural network approach, is that one can even use small data of a few hundreds to build the system. Further, its performance can be improved over time as it learns more about the future. We have tested the feasibility of the system using citizens’ affective responses to CO₂ storage and the system demonstrated 90.476% performance.

1. Introduction

Social acceptance of energy projects is the social license or social permission granted to an energy project and its infrastructure by stakeholders at the market, socio-political and community levels [1,2]. It is not a one-time approval of the project. Its nature is dynamic and evolves over time. It means that the fact that an energy project has been approved initially does not suggest that it has achieved its social license and will be acceptable forever. Overtime, the social license can be withdrawn if the stakeholders feel a need to do so [3,4]. The withdrawal can manifest in different forms, including protest action and vandalism such as the case in the Niger Delta on oil and gas pipelines [5]. Literature reports from wind energy, CO₂ capture and storage and nuclear energy show that a lack of social acceptance for an energy project can bring the project to a stop or delay the policy decisions for its implementation if overlooked in the decision-making process. The retaining and withdrawal of the social license is influenced by a number of psychological factors. These include the stakeholders’ emotion, competence and integrity-based trust of project managers, moral values and risk and benefits perception. These psychological factors are often seen by some project developers as non-technical issues and are often overlooked in the decision-making process [6,7,8,9]. The work of [7] even found that they sometimes labeled the stakeholders as too emotional and irrational. But the author’s findings add that the project managers’ decisions are also without emotion. Due to this emotionally influenced decision, even if they are aware of the inherent risk of the energy project if developed and deployed, they may overlook it. The public is not so naïve and are aware of some of these biases inherent in the top-down decision-making process in energy projects. These days, the withdrawal of the social license of some CCS, wind energy and nuclear energy projects in some countries has made project developers and decision-makers promote these energy technologies to realize that they can no longer take this emotional side of their stakeholders for granted and should rather address it [7].

To address this problem, in recent years, the recommended practices of developers for engaging citizens and their network of stakeholders to gain insights into their perceptions, sentiments and emotions have exclusively relied on mainstream consultation techniques and large-scale deliberative engagement [10]. These standard approaches have merit, but their main limitations include social desirability bias and lack of communication after the workshop which makes the state of the social license uncertain. Further, over time, they could become tedious due to repetitive tasks. The approach is also resource-intensive and time-consuming and may even contribute to logistic emissions (see Discussion for more review). One approach of overcoming this limitation is harnessing the power of artificial intelligence and machine learning to construct decision-making algorithms to aid the human experts [11]. To demonstrate this approach, in our previous work, we provided a framework for modeling emotional behavior in the energy system. What is missing in this work is how to adapt the framework to the communication and engagement problem discussed above to overcome some of the limitations above. Building on this prior work [11], in this present study, an artificial intelligence-powered communication and engagement agent to complement the standard engagement practice is proposed and tested.

This AI-powered engagement agent allows the project developers and decision-makers to work with artificial intelligent agents using deep machine learning to interact with their stakeholders and gain insights into the future. Deep machine learning, also known as deep learning, is a way to train computers to do what comes naturally to humans: learn from examples. It can model complex structures in data and develop the capability to enable decision-makers to make accurate predictions of the future with limited human influence [12,13]. Using this deep learning technique as shown in Figure 1, an AI agent powered by deep learning is an intelligent agent that perceives its environment using the neural network’s capability and reacts. Its smartness can be quantified as predictive accuracy. Predictive accuracy is how well an AI agent performs the intended action relative to the reality [14,15].

Human agents have sensory organs to get information (perceive) from the world (environment) and have muscles (effectors) to take actions in response to the percept. An intelligent agent, on the other hand, can be a robot agent (Figure 1) or a software agent. The proposed communication and engagement agent for an energy project is a software agent. A software agent deployed as a chatbot is tasked to perform specific tasks for a user and possesses a degree of intelligence that permits it to perform part of its task.

To demonstrate the practical feasibility of the proposed method, CO₂ capture and storage (CCS) was used as a case study. It is interesting for the study because it is among the portfolio of low-carbon energy technologies that have emerged without challenges in social acceptance, including nuclear energy, wind energy and hydrogen technology. Despite its ability for radical emission reduction to contribute to fighting climate change [16], in some countries such as the Netherlands, some CCS demonstration plants have been brought to a stop. This is because of a lack of social acceptance in addition to other challenges such as costs [6].

The paper is organized as follows: In Section 2, the theoretical framework of the proposed method is presented. In Section 3, a simulation experiment demonstrating how it can be applied is presented. In Section 4, the implication of the method is discussed with concluding remarks.

2. The Framework of the Proposed AI Communication and Engagement Agent

2.1. General Description of the System

Building on the work of [11] and using a CO₂ capture and storage (CCS) project as a case, Figure 2 presents the block architecture of the proposed AI communication and engagement agent in the CCS context. In the example, it is designed to help the project managers gain an understating of their stakeholders’ emotional behavior to predict their feelings about the energy project as it moves through different life cycle phases.

In real life, the system will be implemented as a chatbot. As a real-time agent, it will offer 24/7 services to the citizens faced with the energy project in question. As shown in the block architecture, the citizens will be able to chat with the agent through voice or text, via Messenger or WhatsApp. As indicated in Figure 2, the system has two agents that communicate with one another to engage the citizens. Agent 1 is the front-end agent and Agent 2 is the back-end agent. In real life, Agent 1 can be implemented via conversational platforms such as Dialogflow. Dialogflow is a Google software that helps to build conversational applications in various languages and on multiple platforms, including Messenger and WhatsApp. It uses a pre-trained natural language processing algorithm that runs through Google servers (cloud). Agent 2 can be implemented using the Keras deep learning library with the Google Tensorflow back-end.

In the real-time communication and engagement process, Agent 1 handles the basic small talk (e.g., How are you? How are you feeling about the project?). It also includes the domain questions on the energy project in question that the project managers want to ask the citizens to gain insights into their feelings about the project. Using these two agents, the system works as follows when deployed: (a) the user opens the communication and engagement chatbot (Agent 1) via the Dialogflow platform; (b) the user chats with Agent 1; (c) Agent 1 collects user answers and stores them on a database; and (d) after the user submits the last answer, Agent 1 will tell the user, “Please, wait a moment, I am processing your information.” Whilst the user is waiting, Agent 1 seeks advice from Agent 2. Therefore, behind the scenes, it is not Agent 1 that is making the final decision. It is Agent 2 that takes the answers from Agent 1 as input, $X$, processes them with the three inference algorithms (receptor, DNN and effector algorithms) and makes the prediction. This prediction is then passed on to Agent 1 in real time for it to communicate the answer to the user. It then gets feedback and communicates this to the project developers for the human cognitive task to enable appropriate intervention. In this paper, we will focus on Agent 2 since it is responsible for the decision-making task. Survey responses will be used to replicate Agent 1’s task.

2.2. Basic Mathematics behind the System

In reference to Figure 2 when a user interacts with the chatbot, Agent 1 takes the stakeholder responses as $X$ and communicates this input to Agent 2 for further processing. In this response, $X$ is the features that the literature underlined that could predict user action towards an energy project [17,18]. As indicated in Figure 2, examples of some of the features are Trust in the competence of actors denoted as $X_1$ and Trust in the integrity of actors, $X_n$. This trust feature collects information on how the stakeholders appraise the competence and integrity of the actors responsible for the implementation and management of the energy project [18]. Assuming the project is yet to be implemented, in line with the works of [17,18], the chatbot can ask the user; “Do you trust in the project developers to take good decisions about storage of CO₂ in your area?” Assuming in response the user wrote “I trust them very much,” this $X_1$ information from Agent 1 will be received by Agent 2. Agent 2 will then process this information using three subsystems. These three subsystems in the inference engine of Agent 2 mimic the behavior of the receptors, neurons and effectors in a biological neural system as indicated in Figure 2. How these three subsystems communicate to each other to arrive at a conclusion on a user feeling is inspired by the principles of stacked generalization [19]. Stacked generalization is an ensemble method that combines the prediction of different learning algorithms into one using a meta-algorithm [19,20].

Inspired by this principle, mathematically, when Agent 2 receives the input $X$ features from Agent 1 via the chatbot, it reconstructs this information into new representations called fuzzy Likert representations, $X^{FL}$. Computationally, it uses the mathematics of fuzzy logic combined with Likert measurement. The combination is called fuzzy Likert [21,22]. The combination enables the responses to be measured within a Likert scale of 0 and 1, but the distance between responses on the scale is known unlike the traditional Likert. In this scale, “0” is the minimum weight (e.g., I don’t trust them at all) and “1” is the maximum weight for responses (e.g., I trust them very much). Fuzzy logic accommodates the degree of partial truth. Therefore, unlike Boolean logic, any responses between 0 and 1 (e.g., I somehow trust them) can also be accommodated by the system if a user expresses them and assigns them weight (for review about fuzzy mathematics, see [23,24].

The fuzzy Likert system uses a fuzzy logic inference engine called the Takagi–Sugeno–Kang (TSK) fuzzy model. This fuzzy type allows human experts to tell the algorithm how to perform the $X$ to $X^{FL}$ mappings. However, the TSK fuzzy Likert system can be optimized using the neural network’s capability to adapt the system to a more complex response. When this is done, it reduces the number of rules which need to be defined by the human experts in the loop. The human experts can teach the system using IF (Premise)…THEN (Conclusion) rules. In this computation, a fuzzy rule $R_s$ is represented as

$$R_s:\mathrm{If}{X}_1\mathrm{is}{A}_1^s\mathrm{and}\dots X_n\mathrm{is}{A}_n^s\mathrm{then}\mathrm{Y}\mathrm{is}{A}_y$$

(1)

In reference to Equation (1), the fuzzy rule output $\mathrm{Y}$ is a polynomial function of the inputs, and the rule is then expressed as

$$R_s:\mathrm{If}{X}_1\mathrm{is}{A}_1^s\mathrm{and}\dots X_n\mathrm{is}{A}_n^s\mathrm{then}$$

$$\mathrm{Y}=a_{s1}{X}_1+\dots +a_{sn}{X}_n+b_s$$

(2)

The defuzzification of the final output calculates the weighted consequent value of a given rule as

$$y^L=\frac{{\sum }_{R_s\in ℛℬ}{h}_s^I P_{s }\left(x_1^l, \dots ‥x_{n }^I\right)}{{\sum }_{R_s\in ℛℬ}{h}_s^I }=\frac{{\sum }_{R_s\in ℛℬ}{h}_s^I v_s^l }{{\sum }_{R_s\in ℛℬ}{h}_s^I }$$

(3)

RB is the rule base of a given set and $R_s\in ℛℬ.e_1\left(x_1^l\dots ‥(x_n^I\right)$ is the input; where $h_s^I=T\left(\mu A_1^s\left(x_1^l\right),\dots .\mu A_n^s(x_n^I\right))$ is the compatibility degree of the instance $e_1$ with the rule $R_s;T$ is a T-norm; and $v_s^l=P_s\left(x_1^l,\dots ‥x_n^I\right)=a_{s1}{x}_1^l+\dots .+a_{sn}{x}_n^I+b_s)$ is the output of $R_s$ and takes the values from $e_1$ (Cozar et al., 2017).

Using this inference method, depending on the complexity of the problem, the fuzzy Likert inference engine can have the following five linguistic rules $R_s=\left(R_1\dots R_{5 }\right)$ to convert the $X \mathrm{variables}$ to $X^{FL}$. The rules are as follows:

$R_1\mathrm{If}$ the Likert response is VERY LOW = 1, then the output is VERY LOW→[0];

$R_2\mathrm{If}$ the Likert response is VERY HIGH = 5, then the output is VERY HIGH→[1];

$R_3\mathrm{If}$ the Likert response is MEDIUM = 3, then the output is MEDUIM→[0.5];

$R_4 \mathrm{If}$ the Likert response is LOW = 2, then the output is LOW→[0.25];

$R_5 \mathrm{If}$ the Likert response is HIGH = 4, then the output is HIGH→[0.75].

As interval details between the ordinal points on the fuzzy Likert scale are known, the transformation from $X$ to $X^{FL}$ provides a high-dimensional behavioral space. In this high-dimensional space, data augmentation can be performed to obtain augmented big data to train a deep neural network. This data augmentation strategy uses the fuzzy logic-based rules in Equation (3) to mimic image data augmentation, as illustrated in Figure 3.

As indicated in Figure 3, the $X$ feature used to observe the social context of the energy system could be thought of as an image, for example, an original butterfly image. As indicated in Figure 3, in the image data problem, after data augmentation, we obtained augmented forms of the images such as de-texturized, de-colored and flipped. The $X^{FL}$ features labeled $X_1^{FL}\dots \dots .X_n^{FL}$ after the data manipulation could be thought of as the augmented images. These $X_1^{FL}\dots \dots .X_n^{FL}$ are the data representations that are extracted from the high-dimensional data space of $X^{FL}$ features of each predictor in the original datasets to train the fuzzy-driven deep neural network (DNN) algorithm shown in Figure 4. The simulation experiment in Section 3 shows this example.

2.2.1. The Algorithm’s Output Decision Using Deep Learning

After the fuzzy Likert system has accomplished its task, the $X^{FL}$ features are passed onto the next subsystems (ensemble fuzzy deep neural network, DNN algorithm). The system mimics the behavior of the biological neurons in the human brain. Ensemble in this work means that the decision is made by different DNN models and their decisions are combined. In real life, the fuzzy-driven deep neural network can be implemented using Keras with the Tensorflow back-end. It is trained using the $X^{FL}$ features. The $X^{FL}$ features are very important in the architecture, especially when the data at hand are small. In a situation where one has big data, one can decide to by-pass step 1 and directly train the system with $X$ features. In this case, the interpretation of the model will be lost. This makes $X^{FL}$ features important features in the system architecture even if one has big data at hand.

Using the trust in actors example, $X_1\dots .X_n$ in Figure 2, let $X_{trust}$ represent the $X^{FL}$ feature for trust in the competence (CBT) of the project actors,$X_1$, and integrity (IBT)-based trust be $X_n$ after the data transformation. After the data transformation, the fuzzy deep neural network receives the $X_{trust}$ information including other responses on other predictors as input, as illustrated in Figure 4.

During computation, the incoming $X^{FL}$ features are multiplied by an appropriate weight function, $w$, and then summed up. The result is recalculated by an activation function, $f$, plus a bias (+1). Mathematically, the output decision of a neuron in the network is expressed in Equation (4):

$$Affective feelings, y=f({\sum }_{i=1}^P{X}_{trust}{W}_1+bias)$$

(4)

From Equation (4), in the decision-making process as shown in Figure 3, the neurons in the deep neural network construct its hypothesis on the problem by learning directly from the $X^{FL}$ features used to train the network. In mainstream practice, human experts would have defined the hypothesis, but it is carried out otherwise in this context. The relationship between features and labels is built out of simpler ones to form a graph of hierarchies. As shown in Figure 3, these graphs are many artificial neurons which are connected layer to layer. In this connection, an output of a neuron automatically becomes input information to another [11,12,25,26]. In the learning process, the network is trained using supervised learning. In supervised learning, the network learns to predict the affective feelings on the energy project in question based on example input–output pairs. However, because the system is an ensemble system, different networks are trained to learn to solve the problem and output their own decisions, Y = ($y_1, y_2\dots y_n$). In this decision, $y_1$ is the decision of network 1, $y_2$ is the decision of network 2 and $y_n$ is the decision of the final network.

2.2.2. Reaction of the Algorithm

As indicated in the block architecture in Figure 2, the final decision, $y$, from the fuzzy-driven deep neural network is then passed onto the final algorithm which acts as the biological effectors. It is responsible for translating the decisions from the ensemble fuzzy-driven deep neural network into the final decision. As illustrated in Figure 2, this final decision is communicated to Agent 1 to communicate to the user. The effector algorithm also uses fuzzy rule inference similar to the first subsystem. As indicated in Figure 2, this final decision is computed to Agent 1 as $y_{predicted}$.

It will be recalled that the proposed communication and engagement agent is a conversational agent. The final subsystem, therefore, is not expected to just output a final prediction of the affective feelings of the citizens to the project managers. Instead, the system is expected to use this emotional intelligence to re-engage the citizen in an emotional conversation. The end goal of this conversation is to understand their feelings and communicate the reason behind the user’s feelings about the decision-makers and project developers. The decision is structured around three affective clusters, as shown in Figure 5.

These affective clusters are positive affective feelings (PA), moderate affective feelings (MOD) and negative affective feelings (NA). Using the CO₂ storage example as shown in

Table 1. Structure of the data and how the agent is expected to react to a citizen and collect their qualitative information and communicate it to the project developers for appropriate intervention.

Prediction	Associated Statements (Intended Action of the Agent)	Numerical Fuzzy Class Label
PA (positive affective feelings)	“Oh, my intelligence tells me you don’t have a problem having the CO2 storage close to your house. Did I get your feelings right, and why this decision?”	0 to 0.455
NA (negative affective feelings)	“Oh, my intelligence tells me you seem to have a problem having the CO2 storage close to your house and want it far away. Did I get your feelings right, and why this decision?”	0.645 to 1
MOD (moderate feelings)	“Oh, my intelligence tells me you seem to have a problem having the CO2 storage close to your house and want it far away. Did I get your feelings right and why this decision?”	0.455 to 0.645

, PA indicates positive affective feelings toward the CO₂ storage. NA indicates negative affective feelings. It is a prediction for those who have problems with the CO₂ storage being close to them and who are emotionally averse to its approval. Finally, MOD indicates moderate feelings. It is a prediction for those with indecisive feelings and struggling to take a position.

Within this three-part affective framework in Figure 4, as indicated in Table 1, if Agent 2 predicts the final decision as PA, Agent 1 will say to the user, “Oh, my intelligence tells me you don’t have a problem having the CO₂ storage close to your house. Did I get your feelings right, and why this decision?” If Agent 2 predicts the final decision as NA, Agent 1 will say to the user, “Oh, my intelligence tells me you seem to have a problem having the CO₂ storage close to your house and want it far away. Did I get your feelings right, and why this decision?” Finally, if Agent 2 predicts the final decision as MOD, Agent 1 will say to the user, “Oh, my intelligence tells me you seem to have a problem having the CO₂ storage close to your house and want it far away. Did I get your feelings right, and why this decision?” These scripts are designed depending on the need of the project developers. In this context, a typical fuzzy IF…THEN rule in the inference engine of the effector algorithm will be as follows: IF Agent 2’s prediction is PA THEN Agent 1 says to the user “Oh, my intelligence tells me you don’t have a problem having the CO₂ storage close to your house. Did I get your feelings right, and why this decision?”

This communication will likely encourage the user to freely express his or her feelings qualitatively. The system will then communicate this qualitative information as feedback to the project developers. However, that part of the system design is beyond the scope of this work. This qualitative information will give them insight into the problem. This could assist them to devise the intervention needed to address people’s concerns. In the section that follows, an evaluation of the model is presented using a hypothetical CO₂ storage.

3. Materials and Methods

3.1. Simulation Experiment and Assumption

The experiment was conducted using a hypothetical CO₂ storage. It was conducted with the assumption that the project was yet to be proposed for the citizens’ vicinity. It was assumed that the CO₂ storage would be proposed near their homes. As highlighted in the theoretical framework, the inference engine of Agent 2 was the main focus, so we will use a survey with questions and answers to represent Agent 1 and feed into the inference engine of Agent 2. The goal of this section is to present a detailed description of the simulation experiment.

3.2. Description of the Dataset and Measurements

To obtain sample data to develop and test the system, 198 volunteers from 15 countries (both developed and undeveloped) were engaged in the CCS discourse on a hypothetical CCS project. They are heterogeneous and come from different economic backgrounds. They range from unemployed to medical doctors, lawyers, university lecturers, environmentalists, teachers, PhD students, secondary school graduates, health professionals, etc. The youngest respondent in our sample was 18 years and the oldest respondent was aged 70.

The media used to reach these respondents were LinkedIn, Facebook, WhatsApp and friend referrals. Some of the respondents were also our Bachelor degree students in Finland. The participants received both video and written information. This content came from both proponents’ and opponents’ materials so that they were not exposed to one-sided information. Some of the proponents’ and opponents’ materials used included the video content on YouTube of Scottish Power Professor of Carbon Capture and Storage Professor Stuart Haszeldine titled “Fuelling the Future: Electricity with Carbon Capture and Geological Storage,” the video content on YouTube of the Zero Emissions Platform (ZEP) titled “The Hard Facts behind Carbon Capture and Storage” and the video content on YouTube of Science TV presenter and climate change communications specialist Yasmin Bushby titled “Carbon Capture & Storage.” Yasmin Bushby’s material takes a neutral viewpoint in explaining CCS and its advantages and challenges in plain language to a lay audience. The Greenpeace material titled “Carbon Capture Scam” was also provided to the volunteers. The YouTube video content of the carbon capture demonstration plant Technology Center Mongstad (TCM) in Norway was added to give the participants a feel of how the plant might look like in real life in their local area. Since they had different materials to choose from, this limited our influence on the information shaping their attitudes and perceptions. The questions were scripted in a way that encouraged them to imagine that the project was taking place where they lived, in line with [27,28]. For example, instead of asking them to rate “trust in government,” we said “trust in the government of my country.” This was purposely done to replicate local context information.

They were observed using five psychological constructs (Trust in actors, Risk perception, Benefit perception, Reaction to Proximity and Affective feelings toward the CO₂ storage) in line with Figure 4. These five constructs are associated with 25 input predictors that predict citizen responses to a CO₂ storage according to the literature [18,27,28,29,30].

Table 2. Predictors used in the simulation experiment to predict the emotional response to the CO₂ storage nearby.

IV	Input Variables	IV	Input Variables (IP) and Output Variables (OV)
IV1	Competence-based trust (CBT): government of my country	IV14	Risk perception: bad effects on trees and plants by sudden leakage of CO2
IV2	CBT: industry (e.g., utility companies, oil and gas companies)	IV15	Risk perception: bad effects on human health from leaked CO2
IV3	CBT: environmental non-governmental organizations (NGOs)	IV16	Risk perception: pipeline being destroyed by earthquake
IV4	CBT: Environmental Protection Agency (EPA) of my country	IV17	Risk perception: bad effects on soil from leaked CO2
IV5	CBT: scientists and engineers in my country	IV18	Risk perception: acidification of seawater by leaked CO2
IV6	Integrity-based trust (IBT): government of my country	IV19	Risk perception: pipeline being destroyed by corrosion
IV7	IBT: industry (e.g., utility companies, oil and gas companies)	IV20	Risk perception: reservoir containing the CO2 being destroyed by an earthquake
IV8	IBT: environmental non-governmental organizations (NGOs)	IV21	Benefit perception: myself
IV9	IBT: Environmental Protection Agency (EPA) of my country	IV22	Benefit perception: my family
IV10	IBT: scientists and engineers in my country	IV23	Benefit perception: my future children yet unborn
IV11	Trust in actors as a team	IV24	Benefit perception: the environment
IV12	Overall risk perception towards CCS	IV25	Reaction to proximity of the CO2 storage
IV13	Risk perception: sudden release of a large amount of stored CO2	OV26	Affective response to the CO2 storage nearby (output predictor of which the algorithm built its emotion intelligence from)

presents the predictors.

Using standard questionnaires adapted from the work of [27,28,29], the appraisal of these predictors was captured using a five-point Likert scale. Their emotions were captured as an affective reaction in line with [27,28,30,31,32,33]. Having obtained the raw data, the simulation experiment proceeded to data pre-processing.

3.3. Data Pre-Processing and Training

As illustrated in Figure 6, the intelligent agent learns from examples to extract a hypothesis and generalize it to unseen cases through optimization and testing.

Due to this learning procedure, the 198 raw datasets, $X$, were divided into a training dataset and testing dataset. In machine learning, the training dataset is used for constructing the inference model (in the case of this paper, Agent 2). During training and validation, the data can leak to the model. The testing data are therefore used to offer an objective evaluation. The test data help in quantifying the predictive accuracy of the model to estimate its ability to generalize to unseen cases. To achieve this goal, the 198 raw data were randomly divided into a training dataset and testing dataset using the 60/40 rule. This split should have led to a 118.8 dataset for training and 79.2 dataset for testing, but a subjective decision was made. This subjective decision rounded up the decimal split and it led to a 114 (raw training data) and 84 (raw testing data) split. This decision was made to ensure that the training dataset was not dominated by, for example, responses from the participants from the developed or developing world. If it is overlooked, it would have increased the chance of the model becoming biased since it might learn one-sided information.

The data were then pre-processed. All missing information was filled with the global constant, “I don’t know.” Having acquired these values after pre-processing, the receptor algorithm (TSK fuzzy Likert inference systems) rule strategy in Section 2.2 was applied to the raw dataset. Five (5) fuzzy IF...THEN rules were defined as human-level rules to initiate the learning of the receptor algorithm, as shown in

Table 3. Sample fuzzy rules used in the experiment to transform the original data $X$ on a Likert scale to their corresponding fuzzy Likert features, $X^{FL}$.

Rules	IF ... $\mathit{X}$	THEN ... ${\mathit{X}}^{\mathit{F}\mathit{L}}$	Fuzzy Scale Range of ${\mathit{X}}^{\mathit{F}\mathit{L}}$ on a 5-Level Membership Function, ${\mathit{\mu }}_{\left(\mathit{x}\right)}$
Rule 1	LOW	LOW	0 to 0.134
Rule 2	SOMEHOW LOW	SOMEHOW LOW	0.134 to 0.44
Rule 3	MEDIUM	MEDIUM	0.44 to 0.644
Rule 4	SOMEHOW HIGH	SOMEHOW HIGH	0.644 to 0.9444
Rule 5	HIGH	HIGH	0.944 to 1

.

This helped to obtain the corresponding fuzzy Likert features, $X^{FL}$. The 114 training data would have led to overfitting when used to train the second subsystem (DNN) because it is a complex algorithm. If the model overfits, it can perform excellently on the validation dataset as part of the training, but when it is tested with the test data, it can lead to poor generalization [13,14,34]. To prevent this, data augmentation was performed on the high-dimensional data space of the $X^{FL}$ labels of the 114 training data. Through this data augmentation, a large experimental dataset of 72,105 training samples was collected. This augmented dataset using the $X^{FL}$ features and lables was then used to train three deep neural networks. The training was implemented using the Keras machine learning library with the Google TensorFlow back-end. The following is the architecture of the ensemble DNN of Agent 2: training iterations: 200 epochs; learner type: neural networks; activation function: rectified linear unit (ReLU); optimization algorithm: stochastic gradient descent; learning rate: 0.003; regularization: dropout; loss function: categorical cross-entropy; number of output nodes: 3 classes (PA, NA and MOD); hidden layer: Model 1 is a 12-layer network (including input and output layer, Model 1); Model 2 is an 11-layer network; and Model 3 had 10 hidden layers.

3.4. Agent Validation and Testing

To objectively evaluate the model, the 84 out-of-sample data (test data) were presented to the ensemble trained models. The models’ decisions were then used as input to the final subsystem (effector algorithm) (Figure 3). Similar to the first subsystem, the final subsystem was implemented using the MATLAB Fuzzy Logic Toolbox. The predicted values were compared to the self-reported affective feelings of the 84 volunteers (test data).

Table 4. Result (performance) from the CCS communication and engagement agent (Agent 2) to be shared with Agent 1 to communicate with the user using its corresponding human languages in Table 1.

Decisive Cases (Positive Affective, PA Feelings, and Negative Affective, NA Reactions)		Indecisive Cases (Moderate Feelings, MOD)
Number of cases: 76		Number of cases: 8
Correct	Wrong	Correct	Wrong
70 cases	6 cases	6 cases	2 cases
Overall performance on the 84 test data		The algorithm made 8 mistakes and predicted 76 correctly, amounting to a 9.523% error with approximately 90.476% predictive accuracy.

presents the simulation results on the test data.

4. Discussion and Implications

In this paper, we contributed to advancing the knowledge development of the state of the art of applications of artificial intelligence (AI) in energy systems with a focus on the phenomenon of social acceptance of energy projects. In this contribution, an alternative approach for complementing the mainstream communication and engagement practice in an energy system is proposed and tested. As briefly stated in the Introduction, in the mainstream approach to engagement, the citizens are engaged through face-to-face interactions through workshop and other outreach programs. The strength of these mainstream approaches is that they create a face-to-face platform for the project managers and the stakeholders to influence each other with their worldviews. This helps the project managers to gain insights into the perceptions, sentiments, emotional states, etc., of their stakeholders on their operations. It is a good forum for the stakeholders to influence the project outcome and co-create a socially acceptable energy project. At the end of the workshop, the project developers will have an idea if the stakeholders are willing to grant them their social license or not. This information helps them to devise interventions and address key concerns that emerged.

This social science approach to engagement has worked well for many energy projects including wind, nuclear and CCS projects. The limitation, however, is that after the workshop, there is a break in communication. This break in communication makes the future acceptability of the project uncertain. It is until the next workshop is organized that the project managers will have a chance to evaluate the state of the social license initially administered to the energy project in question. It is arguable to think that the break in communication can take some projects by surprise when the state of the social license, for example, changes from positive to negative as oppose to what was revealed in the workshop. This assertion is well supported in the emotional literature (e.g., see [35,36,37,38]). In this body of work, it is explained that the state of the social license for the project can change because humans are good at predicting their present emotions and perceptions about emotional stimulus events. However, humans are poor at predicting their future emotions on the same stimulus events that were, for example, initially appraised to be positive or negative. In relation to the acceptability of energy projects, it means that if, for example, at the workshop it was learnt that if events A or B change the people will continue to grant their social support, it does not mean that it will be the case forever due to the fragile nature of the social license [4]. The changing behavior is also due to the stakeholders’ knowledge deficit of the future events that may unfold but were unknown to them at the time of decision-making at the workshop. Another limitation is social desirability bias. In the face-to-face workshop, social desirability bias may cause some stakeholders to be dishonest with their true feelings about the energy project which would make the state of the social license uncertain [16].

As highlighted in the Introduction, one approach of overcoming these weaknesses in the mainstream approach is through affective forecasting based on affective computing. This approach does not rely entirely on the human experts’ judgment based on the outcome of the workshop to predict the future. Instead, it brings together the cognitive capability of the human experts’ judgments together with the probabilistic capability of artificial intelligence and the deep machine algorithm. The combined capabilities lead to the development of a machine learning capability to gain better foresight into the future to make accurate predictions on the emotional behavior and assess its influential variables. In this paper, we provided a framework for constructing this kind of AI-powered engagement algorithm to complement the traditional practice and address its weakness. As demonstrated in the simulation experiment, the algorithm was able to automatically detect the affective state of the stakeholders and performed the required engagement task in Table 1 with about 90% accuracy (Table 4). Given the complexity of human behavior, this performance suggests that such an approach is feasible and future work can further develop the idea to facilitate stakeholders’ decisions on energy projects. Relative to workshops, it is highly cost-effective and reduces logistic emissions. This merit stemmed from the fact that it does not need the physical presence of the stakeholders. It takes advantage of the growing penetration of smartphones and the internet which the stakeholders may have already.

5. Conclusions and Limitations of the Work

The objective of this research was to propose an AI-powered communication and engagement agent as an alternative solution to overcome the limitations in the traditional communication and engagement strategy for energy projects. In the methodological section, the proposed tool was presented and tested in the context of CCS. The feasibility of the proposed tool and its implication for the traditional communication and engagement tool were discussed in the preceding section.

In light of the foregoing discussion, in conclusion, in addition to the novelty of the proposed method in relation to the traditional engagement method mentioned in the Discussion, it is also worth mentioning the following. The proposed method is not a static system which always needs human intervention to understand the changing future. It is learnable and adaptable. It means that the more it interacts with the stakeholders, the more it becomes intelligent and self-adaptable to the changing future behavior. Of course human experts need to guide the algorithm during learning but how it learns is beyond human influence. Further, the proposed system can be implemented in practice without limitation to the big data requirement unlike the mainstream deep learning algorithm. This is what made it possible for us to use only 198 datasets in our simulation experiment. Despite the strengths of the proposed communication and engagement agent, its main limitation is that human interaction is still missing when engaging with people. Even though it is a conversational agent that behaves like humans, it can never replace the human dimension in the standard engagement approaches. It is therefore recommended that the proposed system be seen as complementary to the standard method and not a replacement.

Another limitation is that the AI-powered communication and engagement tool is emotionless in nature. However, the advantage of this emotionless nature of the algorithm in relation to human ethics in the traditional approach is that it neither knows a proponent or an opponent during the engagement. It means that it favors no one. That is to say, when it is implemented with transparency without human influence or brings the human biases into the system reasoning, its decision is objective, especially when the system is well trained. It reports what it sees in the stakeholders’ behavior on the energy project and not what the proponent or opponent expect to see. Ethically, this may be preferable for objective decision-making. This is because studies have shown that research projects in recent years have turned from a research-driven model to a customer-oriented model where the funders define what they expect. It means that if, for example, a proponent funder of a particular energy technology sponsors a research engagement workshop to solicit views about an energy project they support, it is likely that the funder expects support. On the other hand, if an opponent funder sponsors a project, the goal is to do everything possible to convince people to reject it. Evidence of these human biases in promoting one energy technology over the other can be seen in the proponents’ and opponents’ materials used in collecting data for the simulation experiments in the CCS context. In this context, in the data handling, it may happen that the project managers may pay attention to areas of the datasets that support their agenda. However, such human biases are free from the proposed emotionless AI engagement tool when processing the data at hand.