Twitter’s pulse on hydrogen energy in 280 characters: a data perspective

Table 1 shows the list of keywords from multiple languages that are used to extract the data from Twitter. These keywords are selected to comprehend public discourse on hydrogen energy by delving into the viewpoints and global perspectives. To ensure comprehensive coverage, specific keywords were selected in English and Japanese, as these two languages encompass most of such tweets. According to the study (Moernaut et al. 2022) analysing 118 billion tweets between 2009 and 2019, English is the primary language on Twitter, representing 35.6% of the tweets, followed by Japanese with 17.8% of the tweets.

Apart from having the second highest number of Japanese tweets on Twitter, it tops the competitiveness of hydrogen technologies since 2011, as per Astamuse’s ranking. This is attributed to Japan’s superior fuel cell patents, powering hydrogen-led applications in factories, homes, and automobiles (Alshaabi et al. 2021). Japan’s distinction as the first nation to create a national hydrogen strategy in 2017 (Japan: First nation to form national hydrogen strategy 2017) justifies its inclusion in our study. However, the rapid advancements of China in hydrogen-related technology pose a challenge, but we could not include Chinese keywords due to Twitter’s unavailability in China.

South Korea’s significant contributions to hydrogen technology, including a third of the world’s installed utility-scale fuel cells and pioneering commercial fuel cell vehicles, highlight its commitment to hydrogen advancements. Including the Korean language in our study is essential for a comprehensive analysis (The hydrogen economy South Korea: Market intelligence report2021).

We have also included Hindi in our study. India has the third-highest number of Twitter users (23.6 million) (Leading countries based on number of twitter users as of January 2022), and Hindi is the most spoken language in India, with over 584 million speakers. Additionally, India is a major player in the global hydrogen energy market with the Indian government launching the National Green Hydrogen Mission (National green hydrogen mission 2020) and Indian companies heavily investing in hydrogen energy (A fully integrated renewable energy ecosystem by reliance. 2023).

We utilized the Academic and Research Twitter API v2 (Twitter api v2. 2023) to download about 30 million (30,758,600) tweets from January 2013 to December 2022, using keywords listed in Table 1. This ten-year period provides a comprehensive dataset to analyse hydrogen-energy-related discussions, capturing key milestones and the evolving landscape of the topic. Including data from 2013 (the launch year of the world’s first commercial fuel cell vehicle (The hydrogen economy South Korea: Market intelligence report 2021) allows us to track the early stages of hydrogen technology development while extending the analysis until 2022 ensures our study reflects up-to-date information and recent advancements.

The downloaded Twitter data was processed using a combination of Python’s inbuilt standard library and user-defined functions to remove the noise and convert it into a representation suitable for analysis. The following preprocessing steps were performed.

The downloaded Twitter data includes a lang attribute indicating the tweet’s language code identified by Twitter. By utilizing keywords from four languages, we increase the likelihood of capturing tweets in those languages as well as others. A tweet in any language with English hashtags can be identified even without specific language keywords. For example, the Hindi tweet भारत में आई पहली हाइड्रोजन कार, परिवहन मंत्री ने की संसद तक की सवारी (i.e. India got its first hydrogen car, transport minister rides till Parliament) contains several English hashtags (#hydrogen #hydrogencar #hydrogenfuel #hydrogenwater #NitinGadkari #toyotahydrogencar #toyotacars #hydrogenenergy). To analyse language distribution in the unique id dataset, we examined two approaches: (1) tweets downloaded using only English keywords, and (2) tweets downloaded using keywords from all four languages (English, Japanese, Korean, and Hindi).

In order to perform spatial analysis, it is essential to identify tweet locations and plot them geographically. The spatial analysis involves examining the geographic locations of unique id tweets and associated metadata to understand various phenomena, including events and trends that occur in specific locations (Shah and Dunn 2019). By combining spatial locations with a temporal dimension, we conduct spatiotemporal analysis.

The Twitter API offers two different methods to obtain user locations.

Using the GeoPy Python library (Twitter locations. 2023; Python geopy library 2023), we mapped most of the user locations from their profiles to corresponding coordinates (latitude/longitude). However, the presence of a value does not guarantee accurate location information, as some values may be noisy or invalid. In cases of invalid locations, the GeoPy library either provides an empty value or occasionally maps to a false location with valid coordinates.

Since only a minimal percentage (0.58%) of tweets contain exact coordinates (i.e. GPS-enabled location), we expanded our coverage by leveraging profile locations. Approximately 61.58% of tweets had some form of value in their profile location field. The GeoPy Python library identified accurate GPS coordinates for approximately 42% (9,189,714) tweets. Notably, most of the tweets’ location is expressed as a country name in their profile and GeoPy provides the approximate centre of the country as a default coordinate. For example, the coordinates 36.57,139.24 correspond to Kuroganecho for Japan, -24.78, 134.75 correspond to Ghan NT for Australia, 22.35,78.67 correspond to Madhya Pradesh for India, and 39.78, -100.44 correspond to Oberlin for the USA.

Utilizing geospatial coordinates and their corresponding frequencies, we utilized the Geopandas Python library (Geopandas python library 2023) along with Matplotlib to generate visual representations of geospatial data on world maps for each year. A world map is created using scatter plots, where a marker represents each data point on a two-dimensional coordinate system (i.e. longitude and latitude of a location).

The tweet object retrieved from the Twitter API includes essential metadata about individual tweets, such as tweet id, creation date, text content, author information, and associated public metrics (e.g. impression count, like count, retweet count, replies, and quote count). The impression count denotes the number of times the tweet has been viewed; the like count indicates the number of users who liked the tweet; the retweet count represents the number of times the tweet has been retweeted; the reply count signifies the number of replies received by the tweet, and the quote count indicates the number of times the tweet has been quoted (i.e. retweeted with a comment). These public metrics are particularly interesting as they provide insights into the reach, engagement, and popularity of tweets within the Twittersphere, shedding light on how well tweets resonate with the Twitter audience. We extracted tweet-level public metrics for each original tweet in our dataset of unique id and unique text tweets.

The user object retrieved from the Twitter API provides valuable metadata about user accounts, including user id, account creation date, user location, user description, verification status, and public metrics like follower count, following count, tweet count, and listed count since the creation of the account. We are particularly interested in the verification status and public metrics, which offer insights into user behaviour and influence on Twitter (2023). To facilitate user-specific analysis, we eliminated duplicate users from both the unique id and unique text tweets and examined the publicly available metrics associated with each user. This analysis involved categorizing users based on different ranges of these metrics to a bin. For example, the bin [0,500] for follower count includes users with followers below 500, including those with no followers. The bin ranges begin from zero and go up to the maximum number observed for each criterion, with the highest follower count recorded in our dataset being 133,996,064.

Given the structure of Twitter’s API v2, hashtags and mentions are not only found within the tweet content, but are also stored in the same entities field under the hashtags attribute and the mentions attribute of a tweet (Twitter api v2. 2023). Considering their shared placement within the tweet object, we processed both hashtags and mentions simultaneously. We processed the unique id tweets dataset by dividing it into yearly segments. We identified the most frequently occurring hashtags and mentions for each year and visualized them on a heat map. We extracted the top 200 hashtags and mentions to gain further insights and visualized them using a word map. Non-English hashtags were translated into English equivalents, distinguished with the suffix _x.

Hashtag analysis reveals trending topics and themes, reflecting the dynamics of public discourse. Mention analysis identifies influential users or organizations and their relationships, offering insights into user engagement and active participation in discussions. The combined analysis of hashtags and mentions provides an understanding of their prominence and occurrence over time.

The analysis of the lang attribute on unique id and unique text tweets reveals that the dataset encompasses a wide range of 72 language codes, including a small proportion that do not correspond to actual spoken languages. These language codes,² namely qam, qct, qht, qme, qst and zxx, signify the tweets that contain only mentions, cashtags, hashtags, media links, very short texts, and media or Twitter card, respectively.

To assess the impact of including keywords from different languages on tweet volumes, we analysed the distribution of tweets per language downloaded with English language keywords only. As shown in Fig. 2, most tweets were posted in English and Japanese in both cases. As seen in Fig. 2a, based on the data collected using English-only keywords, approximately 83.5% of the tweets were in English. Japanese accounted for the second largest portion, comprising 5–6% of the tweets, and other languages like Korean, German, French, and Spanish made up around 4%. Hindi ranked 14th with 0.19% of tweets. As seen in Fig. 2b, using the unique id tweets gathered using keywords from four languages, the top five most expressed languages are Japanese (55.24%), English (37.99%), Korean (2.31%), German (0.60%), and Hindi (0.38%), accounting for 96.53% of hydrogen-related tweets. However, when considering the unique text tweets, English becomes the top language with 47.94% tweets, followed by 43.87% Japanese tweets.

Discussion: It is interesting to note that using language-specific keywords for data extraction does not necessarily limit obtaining the tweets to the same language, as code-mixing can occur where keywords are used alongside words from different languages (Rijhwani et al. 2017). While the data collected using English-only keywords contains predominantly English tweets, Japanese emerged as the predominant language when using keywords in four languages. A closer examination of tweets revealed that some Japanese tweets with different ids shared the same content, potentially indicating the usage of bots for campaigns, advertisements, propaganda, or other purposes (Keller and Klinger 2019). Acknowledging the potential bias introduced into the data by repeated content is important. The lower number of tweets in Hindi may be attributed to a significant portion of Indian tweets being posted in English and other Indian languages (Indian tweets in English 2019). Furthermore, unexpected numbers of German, French, and Spanish tweets warrant further investigation.

Relying solely on English keywords limits the representation of non-English perspectives and fails to capture the nuances specific to different language groups fully. Using multiple language keywords ensures a more inclusive dataset for diverse and global insights on hydrogen energy.

This analysis presents spatial and temporal patterns in tweets and associated metadata. The outcome illustrates the varying levels of tweet activity over time and regions and offers valuable insights into the dynamics of online conversations.

Temporal analysis: Figure 3 depicts the count of tweets and media content over a ten-year period. The heat maps of hashtags, mentions and bigrams, shown in Figures 7 and 9, further reveal how the discussions related to hydrogen and related topics have evolved over time.

Paris Agreement (COP21, December 2015) likely increased global environmental awareness (Paris agreement 2015) and its impact through increased Twitter activities can be seen in Fig. 3. The tweet count steadily increased over time, with a notable spike in May 2016. This spike was primarily driven by Japanese tweets (94.28%), while English tweets accounted for only 4.52%. During this period, the most frequently used hashtags included #Yahooニュース (#Yahoonews), # 水素 (#hydrogen), #niconews, #hydrogenwater, and #fuelcell, among others. This suggests widespread sharing of Yahoo News articles and topics particularly emphasizing hydrogen-related discussions, such as hydrogen water and fuel cells. Several spikes occurred in September 2017, April 2018, September 2018, July 2019, and April 2021. Most tweets were in Japanese, accounting for 34.79, 78%,

58.54, 96.49, and 42.87% respectively. English tweets constituted 54.08, 16.59, 35.29, 3.19, and 53.01% during the same period. Tweets in other languages were minimal, representing only 11.12, 5.37, 6.15, 0.32, and 4.11%, respectively.

In September 2017, the trending hashtags included #hydrogen, #northkorea, #hydrogenbomb, and #fuelcell. Global concerns and discussions on Twitter were sparked by North Korea’s missile tests and their claim of a successful hydrogen bomb test (North Korean nuclear test 2017). Likewise, in April 2018, popular hashtags included #hydrogen, #fuelcell, #水素の音 (the sound of hydrogen), #超会議2018 (#Chokaigi2018) and #超会議コスプレ (#super conference cosplay). In September 2018, hashtags such as #水素 (#hydrogen), #fuelcell, #germany, #energy, #environment, #水素水 (#hydrogen water), #transport, #missionh24, #健康 (#health), #fuelcells, #renewableenergy, and #hydrogentrain were prominent. This indicates a growing interest and discussion surrounding hydrogen energy and related topics. The launch of the world’s first hydrogen-powered train (Coradia iLint) in Germany during that time, built by French TGV-maker Alstom, further highlights the advancements and recognition of hydrogen as a viable and sustainable energy solution (Germany launched world’s first hydrogen-powered train 2018). In July 2019, hashtags such as #fuelcell, #climatechange, #zeroemission, #greenhydrogen appeared. In April 2021, several hashtags related to hydrogen, environmental initiatives, and collaborations between Hyundai and BTS were trending on social media. The popularity of the #hydrogen hashtag could be attributed to increased awareness and discussions around hydrogen as a clean and sustainable energy source. The #hyundaixbts, #hyundai, #bts, #nexo, #btsxhyundai, #nexoxbts, and #ioniqxbts hashtags likely gained traction due to a collaboration or promotional campaign between #Hyundai, a car manufacturer, and BTS, a popular South Korean music group. Additionally, hashtags like #earthday, #fortomorrow, #wewontwait, and #energy were likely trending in connection with Earth Day and discussions about renewable energy and sustainability. The presence of the #トヨタ (#toyota), #greenhydrogen, and #fuelcell suggests Toyota’s involvement.

We also identify and show the count of media files (e.g. images, animated gifs, or videos) posted by users in Fig. 3. To simplify representation, tweets with multiple media files, like images, videos, and animated gifs, are counted as distinct tweets. On average, tweets with media files typically have around 1.41 images, 1.3 videos, and 1.07 animated gifs per tweet. However, when considering all the unique id tweets, the average number of media files decreases significantly to 0.34 images, 0.04 videos, and 0.006 animated gifs per tweet.

Spatial or geospatial analysis: Figure 4 shows the top 10 countries making up 80% tweets. The tweet distribution across top countries is about 33% Japan, 19.3% USA, 9.4% UK, 3.9% India, 3.6% Canada, 3.5% China and 2.9% Australia. The leadership in Hydrogen energy of these top countries is evident through their partnerships and initiatives such as the International Partnership for Hydrogen and Fuel Cells in the Economy (2003), the Hydrogen Council (2017), Hydrogen Europe (2000), and Twitter activities. Note that these countries have a sizable active Twitter user base (January 2022) (Leading countries based on number of twitter users as of January 2022). Surprisingly, despite Korean keywords and presence in the top five languages (Fig. 2(a)), South Korea ranks 11th. This suggests that Japanese and English-speaking communities are more active in hydrogen energy discussions on Twitter compared to the Korean-speaking community.

Global engagement of various nations in hydrogen energy is evident from the world map in Fig. 5. Japan’s highest tweet count (74,639) in 2019 serves as the (maximum) reference for the right-side colourbar. Japan’s sustained dominance from 2013–2022 is noticeable due to its clean energy commitment, pioneering hydrogen strategy in 2017 (Japan: First nation to form national hydrogen strategy 2017), and technology leadership evident by its numerous fuel cell patents (Alshaabi et al. 2021). Including the Japanese language in our dataset may also introduce bias towards Japan.

Figure 5 also reveals that the USA and the UK show increasing tweet activity over time. The USA’s high tweet counts stem from English language inclusion in our dataset and its prominent global hydrogen energy leadership. The USA has an ambitious hydrogen production goal: 10 million metric tons (MMT) of clean hydrogen by 2030, 20 MMT by 2040, and 50 MMT by 2050 (Us national clean hydrogen strategy and roadmap 2021). The USA drives adoption through initiatives like H2USA (H2usa 2013) and California Fuel Cell Partnership (California fuel cell partnership (cafcp) 2013), emphasizing hydrogen technology advancement and commercialization. Similarly, the UK’s global leadership is displayed through government plans like Ten Point Plan (The ten point plan for a green industrial revolution 2020), Hydrogen Strategy for Scotland (Hydrogen strategy for Scotland 2022) and the UK Hydrogen Strategy (Uk hydrogen strategy 2021), promoting hydrogen’s role in sustainable energy vision.

Canada’s notable presence is evident with its 2020 hydrogen strategy (Hydrogen strategy for Canada 2020), aiming for low-carbon hydrogen using abundant natural resources. Canada’s participation in international collaborations such as IPHE (2003) reflects its proactive approach. Europe’s prominence is shown by EU’s Hydrogen Strategy (2020), REPowerEU (2022) plan, targeting 6 GW of hydrogen electrolysers by 2024, 40 GW by 2030, emphasizing cross-border collaboration and international partnerships. Countries like Germany (The national hydrogen strategy—Germany. 2020), the Netherlands (The national hydrogen strategy—Netherlands. 2020), Denmark (The government’s strategy for power-to-x—Denmark. 2021) and Norway (The Norwegian government’s hydrogen strategy 2020) are also actively investing in hydrogen energy.

India and Australia’s growing interest in hydrogen energy can be seen on the world map of tweets. Minimal tweet activities in 2013, but a notable rise from 2016, reflect increased involvement in hydrogen-related discussions. This trend aligns with events such as the Paris Agreement signing in 2015 (Paris agreement 2015) pushing clean energy solutions. Australia’s 2019 National Hydrogen Energy Strategy (Australia’s national hydrogen strategy 2019) spurred interest and increased Twitter activities (Fig. 5g–j). Similarly, India’s 2020 National Green Hydrogen Mission (2020; Iea 2019) boosted Twitter presence (Fig. 5h–j). Figure 5d–j show rising activities, particularly in Japan and the USA (greenish dots). The trend intensified in 2019, with the United Kingdom, India, Australia, and various European countries showing increased engagement (greenish dots and larger blue dots).

To perform the tweets and user-related engagement analysis, various public metrics from the data are discussed in this section to understand the impact, engagement, and interaction levels of tweets.

Metrics impacting twitter engagement: It is notable that a large portion of unique id dataset is original tweets: 44.7% (9,750,520) were original, 41.24% (8,994,728) retweets, 12.5% (2,718,330) replies, and 1.6% (349,282) quoted content. In unique text subset, 58.9% (5,859,413) were original, 10.9% (1,082,852) retweets, 26.8% (2,663,529) replies, and 3.45% (343,190) quoted content.

We analyse only original tweets for insights into unique content. Examination of public metrics of original tweets reveals that each tweet garnered an average of 499 impressions, 173.5 likes, 39 retweets, 20 replies and 7.25 quotes, indicating its potential reach. Most tweets received few responses: 99.45% had ≤ 10 impressions, 94.68% ≤ 10 likes, 97.13% ≤ 10 retweets, 98.96% ≤ 10 quotes, and 98.24% ≤ 10 replies. This follows the power law, where few tweets have high engagement, and the majority have minimal (Chen and Nayak 2011).

User engagement analysis: User engagement analysis in Fig. 6 provides insights into public user metrics like follower count, following count, tweet count, and listed count, revealing diversity in participants. Most users (94.4% for unique id and 92.5% for unique text) have < 5000 followers, while a small fraction (0.07% for unique id and 0.12% for unique text) have > 1 million (Fig. 6a), following the power law distribution. These users with a high number of followers, often referred to as influencers, tend to play a significant role in shaping discussions and exerting influence within the social media landscape.

When examining the following counts, 2.2% of unique id and 2.6% of unique text tweet users follow ≥ 5000 accounts, while the remaining users follow fewer accounts as shown in Fig. 6b. Analysing tweet counts, 48%-50% have posted < 10,000 tweets, while the remaining users tweet aggressively (10,000 to 115 million tweets) as shown in Fig. 6c. Although higher tweet counts may be expected for long-term active users, a count of 115 million tweets is still exceptionally high for any account.

Further analysis of the top 10 user accounts with the highest tweet counts reveals that all of them are Japanese accounts. Some have been active since as early as 2009, while others joined Twitter more recently in 2019–2020. These accounts primarily belong to various organizations, such as retail stores or food chains, where they identify themselves as official accounts for user interaction or promotional campaigns. Few (0.03%-0.05% unique id and unique text) users have > 1 million tweets, rising to 7.43% and 8.09% with counts > 0.1 million.

Listed counts showcase influence, reputation, and reach. Only 0.06% of unique id and 0.12% of unique text users have > 5000 listed counts as shown in Fig. 6d. Distribution of verified and non-verified users reveals that only 1.4% (64,906) users were verified, while a significant 98.6% (4,574,333) were non-verified among unique id users. Similarly, for unique text users, 2.06% (55,406) were verified users and 97.9% (2,632,874) were non-verified users. Verification status does not solely indicate the credibility or quality of the content posted by users. Verification is primarily a means for Twitter to confirm the identity of accounts that are of public interest.

Mention analysis: Another useful aspect of Twitter data analysis is exploring the mention network as it reveals the frequency of a user being mentioned by others in their tweets and reflects the user’s ability to engage in conversations (Cha et al. 2010). Having covered follower and tweet statistics, our focus now shifts to user mentions, offering significant insights into key influencers within the hydrogen energy sector.

Figure 7a presents the mention heat map, showcasing the top 10 mentioned users each year, resulting in a total of 83 mentions over the span of 10 years. Mentions vary in frequencies, for example, hyundai_global being the most used mentions with a frequency of 207,028 and bts_jp_official appearing in top mentions with a frequency of 23,714 in 2020. The heat map displays equal strength for both due to row-wise normalization. This ensures each word’s independent analysis, preventing less frequent mentions from disappearing, despite high positioning on the list each year.

Additionally, Fig. 8a depicts the top 200 mentioned users during the same period. Notably, several users consistently appeared recently and ranked among the top 200 mentioned users of all time across various categories. These categories include industries and industrialists (elonmusk, youtube, hyundai_global, hyundai_japan, toyota, mliebreich, cafcp, uberfacts), politicians (whitehouse, nitin_gadkari, narendramodi), news and information portals (fuelcellsworks, cnn, h2_view, livedoornews), science and scientists (gatapi21, aldiallaboratoy), artists (bts_twt, kugatsu_main) and economists (daitojimari).

As discussed in Sect. 3, we employed multiple subsets of data for this analysis: (1) unique id tweets for hashtag analysis, (2) a subset of unique text tweets containing English, Japanese, Korean, and Hindi tweets for ngram (unigram and bigram) analysis, and (3) an English-only subset of unique text tweets for additional ngram (unigram and bigram) analysis.

The heatmap focuses on showing the top 15 hashtags per year (Fig. 7b), but there are a total of 78 hashtags over the decade, indicating a dynamic and diverse topic landscape. Despite selecting 10 mentions per year, the overall count of hashtags and mentions has remained the same over the decade. For the top 20 bigrams in English-only and all four languages tweets (Fig. 9), we see 83 and 88 bigrams, respectively. This may be because hashtags and bigrams are often reused over time, unlike user mentions associated with a specific hashtag globally.

Non-English words are translated into English in the visualization for clarity. In Figs. 7) and 9, each row represents a word’s yearly evolution. Despite varying frequencies, e.g. #hydrogen (the most used hashtag with a frequency of 267,728 in 2021) and #水素 (#hydrogen) (a Japanese hashtag with a frequency of 12,882 in 2021), both display similar strength in the heat map as shown in Fig. 7b. Similarly, the bigram hydrogen fuel appeared in English tweets in 2022 with a frequency of 45,632, while fossil fuel appeared 9,055 times in the same year as shown in Fig. 9b. The row-wise normalization allows us to analyse each word independently.

Some main findings are as follows: